gpl.txt 0000664 0000764 0000764 00000104513 11340650532 012316 0 ustar quinlan quinlan GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc.
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
Copyright (C)
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see .
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
Copyright (C)
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
.
Makefile 0000644 0000764 0000764 00000002663 11340610156 012432 0 ustar quinlan quinlan #*************************************************************************#
#* *#
#* Makefile for the Cubist system *#
#* ------------------------------ *#
#* *#
#*************************************************************************#
CC = gcc -ffloat-store
OFLAGS = -O3
CFLAGS = -DVerbOpt -g -Wall -O0
LFLAGS = $(S)
SHELL = /bin/csh
# Definitions of file sets
# New file order suggested by gprof
src =\
global.c\
xval.c\
cubist.c\
sort.c\
construct.c\
predict.c\
stats.c\
discr.c\
rules.c\
contin.c\
formrules.c\
formtree.c\
getdata.c\
getnames.c\
implicitatt.c\
instance.c\
modelfiles.c\
prunetree.c\
regress.c\
trees.c\
update.c\
utility.c
obj =\
global.o cubist.o construct.o\
formtree.o prunetree.o stats.o discr.o contin.o\
trees.o\
formrules.o rules.o\
instance.o\
predict.o\
regress.o\
xval.o\
getnames.o getdata.o implicitatt.o sort.o\
modelfiles.o\
update.o utility.o\
all:
make cubist
$(CC) $(LFLAGS) $(OFLAGS) -o summary summary.c -lm
# debug version (including verbosity option)
cubistdbg:\
$(obj) defns.i extern.i text.i Makefile
$(CC) $(CFLAGS) -o cubistdbg $(obj) -lm
# production version
cubist:\
$(src) defns.i text.i Makefile
cat defns.i $(src)\
| egrep -v 'defns.i|extern.i' >cubistgt.c
$(CC) $(LFLAGS) $(OFLAGS) -o cubist cubistgt.c -lm
strip cubist
rm cubistgt.c
$(obj): Makefile defns.i extern.i text.i
.c.o:
$(CC) $(CFLAGS) -c $<
construct.c 0000664 0000764 0000764 00000027471 11501534415 013172 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Routines for constructing rulesets and committees */
/* ------------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#include
/*************************************************************************/
/* */
/* Construct, save and evaluate a single model */
/* */
/*************************************************************************/
void SingleCttee(void)
/* ----------- */
{
FILE *F;
ConstructCttee();
/* Evaluation */
fprintf(Of, T_EvalTrain,
MaxCase+1,
( USEINSTANCES && MaxCase > 20000 ? ", sampled" : "" ));
NotifyStage(EVALTRAIN);
Progress(-(MaxCase+1.0));
EvaluateCttee(Cttee, false);
if ( (F = GetFile(( SAMPLE ? ".data" : ".test" ), "r")) )
{
NotifyStage(READTEST);
Progress(-1.0);
fprintf(Of, "\n");
if ( USEINSTANCES )
{
Free(Case); /* leave actual descriptions */
}
else
{
FreeData(Case);
}
Case = Nil;
GetData(F, false, false);
fprintf(Of, T_EvalTest, MaxCase+1);
NotifyStage(EVALTEST);
Progress(-(MaxCase+1.0));
EvaluateCttee(Cttee, true);
}
else
if ( (F = GetFile(".pred", "r")) )
{
fclose(F);
remove(Fn); /* set by GetFile */
}
FreeCttee(Cttee); Cttee = Nil;
}
/*************************************************************************/
/* */
/* Main construction routines */
/* */
/*************************************************************************/
void ConstructCttee(void)
/* -------------- */
{
int m;
Boolean SaveUSEINSTANCES;
CaseNo i;
double Cases, SumErr=0, Err, FinalErr=0;
/* Preserve original item order */
SaveCase = Alloc(MaxCase+1, DataRec);
memcpy(SaveCase, Case, (MaxCase+1) * sizeof(DataRec));
FindGlobalProperties();
if ( CHOOSEMODE ) USEINSTANCES = true;
/* Set minimum target coverage for a rule as 1% of cases (to a
maximum of 20). However, it must be at least MINSPLIT and
must not exceed cases / MAXRULES */
Cases = MaxCase+1;
MINITEMS = Min(20, rint(Cases / 100));
if ( MINITEMS > Cases / MAXRULES )
{
MINITEMS = Cases / MAXRULES;
}
if ( MINITEMS < MINSPLIT )
{
MINITEMS = MINSPLIT;
}
Cttee = AllocZero(MEMBERS, RRuleSet);
ForEach(m, 0, MEMBERS-1)
{
Cttee[m] = ConstructRuleSet(m);
memcpy(Case, SaveCase, (MaxCase+1) * sizeof(DataRec)); /* restore */
if ( m < MEMBERS-1 )
{
/* Adjust target value for next regression tree */
Err = 0;
ForEach(i, 0, MaxCase)
{
PredVal(Case[i]) = RuleSetPrediction(Cttee[m], Case[i]);
Err += fabs(Class(Case[i]) - PredVal(Case[i]));
Class(Case[i]) = 2 * CVal(Case[i], ClassAtt) - PredVal(Case[i]);
}
SumErr += Err;
}
}
FreeUnlessNil(SaveCase); SaveCase = Nil;
if ( ! XVAL && MEMBERS > 1 )
{
/* Calculate the error reduction achieved by committee model */
SaveUSEINSTANCES = USEINSTANCES;
USEINSTANCES = false;
FindPredictedValues(Cttee, 0, MaxCase);
USEINSTANCES = SaveUSEINSTANCES;
ForEach(i, 0, MaxCase)
{
FinalErr += fabs(Class(Case[i]) - PredVal(Case[i]));
}
ErrReduction = FinalErr / (SumErr / (MEMBERS - 1));
}
/* See whether to use rulesets or rulesets with instances */
if ( USEINSTANCES )
{
MAXD = -1; /* causes InitialiseInstances to set it */
InitialiseInstances(Cttee);
}
if ( CHOOSEMODE )
{
CheckForms(Cttee);
}
if ( ! XVAL ) SaveCommittee(Cttee, ".model");
}
RRuleSet ConstructRuleSet(int ModelNo)
/* ---------------- */
{
RRuleSet RS;
char Msg[20];
CaseNo i;
RuleNo r;
float TempMTSize;
NotifyStage(GROUPDATA);
Progress( -(MaxCase+1.0) );
FormTree(0, MaxCase, 0, &TempMT, Nil);
NotifyStage(ADDMODELS);
Progress( (TempMTSize = (float) -TreeSize(TempMT)) );
AddModels(0, MaxCase, TempMT, Nil);
NotifyStage(SIMPLIFYGROUPS);
Progress( TempMTSize );
Prune(TempMT);
AdjustAllThresholds(TempMT);
Verbosity(1, PrintTree(TempMT, "Model tree"))
NotifyStage(FORMRULES);
Progress( -(MaxCase+1.0) );
/* Restore original target values */
ForEach(i, 0, MaxCase)
{
Class(Case[i]) = CVal(Case[i], ClassAtt);
}
RS = FormRules(TempMT);
ForEach(r, 1, RS->SNRules)
{
RS->SRule[r]->MNo = ModelNo;
}
if ( MEMBERS > 1 )
{
sprintf(Msg, "Model %d:", ModelNo+1);
}
else
{
sprintf(Msg, "Model:");
}
PrintRules(RS, Msg);
FreeTree(TempMT); TempMT = Nil;
return RS;
}
/*************************************************************************/
/* */
/* Check performance of model */
/* */
/*************************************************************************/
void EvaluateCttee(RRuleSet *Cttee, Boolean Details)
/* ------------- */
{
CaseNo i, MaxTest;
double Real, CorCoeff=0, Tests,
AbsErr=0, BaseAbsErr=0, ProdVar,
SumX=0, SumY=0, SumXY=0, SumXX=0, SumYY=0;
if ( MaxCase < 0 ) return;
MaxTest = MaxCase;
if ( USEINSTANCES )
{
if ( Details )
{
/* Copy instances to improve cache performance */
CopyInstances();
}
else
if ( MaxCase >= 2 * EVALSAMPLE )
{
/* For composite models, use only a sample for evaluating large
training sets */
SampleTrainingCases();
MaxTest = EVALSAMPLE-1;
}
}
/* Find the cases' predicted values if they are not already known */
if ( Details || USEINSTANCES || MEMBERS == 1 )
{
FindPredictedValues(Cttee, 0, MaxTest);
}
if ( Details )
{
if ( ! (Pf = GetFile(".pred", "w")) ) Error(NOFILE, Fn, E_ForWrite);
fprintf(Pf, T_Default, Precision+1, GlobalMean);
fprintf(Pf, F_Actual " " F_Predicted " " F_Case "\n"
F_Value " " F_Value "\n"
" -------- --------- " F_UCase "\n");
}
/* Evaluate performance of model */
ForEach(i, 0, MaxTest)
{
Real = Class(Case[i]);
AbsErr += fabs(Real - PredVal(Case[i]));
BaseAbsErr += fabs(Real - GlobalMean);
SumX += Real;
SumXX += Real * Real;
SumY += PredVal(Case[i]);
SumYY += PredVal(Case[i]) * PredVal(Case[i]);
SumXY += Real * PredVal(Case[i]);
if ( Details )
{
fprintf(Pf, "%9.*f %9.*f %s\n",
Precision, Real, Precision+1, PredVal(Case[i]),
CaseLabel(i));
}
Progress(1.0);
}
if ( Details )
{
fclose(Pf);
Pf = 0;
}
Tests = MaxTest+1;
ProdVar = (SumXX - SumX * SumX / Tests) * (SumYY - SumY * SumY / Tests);
if ( ProdVar > 0 )
{
CorCoeff = (SumXY - SumX * SumY / Tests) / sqrt(ProdVar);
}
fprintf(Of,"\n");
fprintf(Of, F_AvErr "%10.*f\n", Precision+1, AbsErr / Tests);
fprintf(Of, F_RelErr "%10.2f\n",
( BaseAbsErr ? AbsErr / BaseAbsErr : 1.0 ));
fprintf(Of, F_CorrCoeff "%10.2f\n", Max(0, CorCoeff));
/* Summarise attribute usage on training data */
if ( ! Details )
{
AttributeUsage();
}
}
/*************************************************************************/
/* */
/* Sample EVALSAMPLE cases from training set */
/* */
/*************************************************************************/
void SampleTrainingCases(void)
/* ------------------- */
{
CaseNo i, j;
double Step;
DataRec xab;
Step = (MaxCase+1) / (double) EVALSAMPLE;
ForEach(i, 1, EVALSAMPLE-1)
{
j = i * Step;
Swap(i, j);
}
}
/*************************************************************************/
/* */
/* Report attribute usage in conditions and models */
/* */
/*************************************************************************/
CaseCount SumCases, *SumCond=0, *SumModel=0;
Boolean *AttUsed=0;
void AttributeUsage(void)
/* -------------- */
{
RRuleSet RS;
int m;
RuleNo r;
Attribute Att, BestAtt;
char U1[5], U2[5];
/* Initialise counts */
SumCases = 0;
SumCond = AllocZero(MaxAtt+1, CaseCount);
SumModel = AllocZero(MaxAtt+1, CaseCount);
AttUsed = Alloc(MaxAtt+1, Boolean);
/* Scan rules in committee */
ForEach(m, 0, MEMBERS-1)
{
RS = Cttee[m];
ForEach(r, 1, RS->SNRules)
{
UpdateUsage(RS->SRule[r]);
}
}
/* Show attributes in order of combined usage */
fprintf(Of, T_AttUsage);
while ( true )
{
BestAtt = 0;
ForEach(Att, 1, MaxAtt)
{
if ( Max(SumCond[Att], SumModel[Att]) >= 0.01 * SumCases &&
( ! BestAtt ||
SumCond[Att] > SumCond[BestAtt] ||
SumCond[Att] >= SumCond[BestAtt] &&
SumModel[Att] > SumModel[BestAtt] ) )
{
BestAtt = Att;
}
}
if ( ! BestAtt ) break;
sprintf(U1, "%3g%%", rint((100.0 * SumCond[BestAtt]) / SumCases));
sprintf(U2, "%3g%%", rint((100.0 * SumModel[BestAtt]) / SumCases));
fprintf(Of, "\t %4s %4s %s\n",
( SumCond[BestAtt] >= 0.01 * SumCases ? U1 : " " ),
( SumModel[BestAtt] >= 0.01 * SumCases ? U2 : " " ),
AttName[BestAtt]);
SumCond[BestAtt] = SumModel[BestAtt] = 0;
}
Free(SumCond);
Free(SumModel);
Free(AttUsed);
}
void UpdateUsage(CRule R)
/* ----------- */
{
Attribute Att;
int d;
SumCases += R->Cover;
/* Attributes used in conditions. Must assemble in table in case
same attribute appears more than once */
memset(AttUsed, false, MaxAtt+1);
ForEach(d, 1, R->Size)
{
NoteUsed(R->Lhs[d]->Tested);
}
ForEach(Att, 1, MaxAtt)
{
if ( AttUsed[Att] ) SumCond[Att] += R->Cover;
}
/* Attributes used in model */
memset(AttUsed, false, MaxAtt+1);
ForEach(Att, 1, MaxAtt)
{
if ( R->Rhs[Att] ) NoteUsed(Att);
}
ForEach(Att, 1, MaxAtt)
{
if ( AttUsed[Att] ) SumModel[Att] += R->Cover;
}
}
void NoteUsed(Attribute Att)
/* -------- */
{
int i;
if ( AttUsed[Att] ) return;
AttUsed[Att] = true;
if ( AttDef[Att] )
{
/* Include attributes that appear in definition */
ForEach(i, 1, AttDefUses[Att][0])
{
NoteUsed(AttDefUses[Att][i]);
}
}
}
contin.c 0000664 0000764 0000764 00000021253 11501534415 012430 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Evaluation of a test on a continuous valued attribute */
/* ----------------------------------------------------- */
/* */
/* Tests on continuous attributes have three possible */
/* outcomes: 1 (N/A), 2 (less than cut), 3 (greater than cut). */
/* This routine finds the best cut for items Fp through Lp and */
/* sets GEnv.Gain[] and GEnv.Bar[]. */
/* */
/* For outcomes 2 and 3, we imagine a least-squares linear model */
/* of the inputs as a function of the continuous attribute. */
/* The residuals of each model have mean 0 and sum of squares */
/* (1-R^2) * N * variance of the inputs where R is the correlation */
/* coeff and N is the number of cases. Note that these models */
/* do not have to be found; all we need is the correl coeffs for */
/* these outcomes. */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
void EvalContinuousAtt(Tree Node, Attribute Att, CaseNo Fp, CaseNo Lp)
/* ----------------- */
{
CaseNo i, Xp, BestI, Cases, Edge, LoK, HiK;
ContValue Val, LowVal, HighVal;
double ThisGain, BestGain=0,
LoSumX=0, LoSumXX=0, LoSumY=0, LoSumYY=0,
LoSumXY=0, LoN, LoMX, LoVX, LoMY, LoVY, LoR,
HiSumX=0, HiSumXX=0, HiSumY=0, HiSumYY=0,
HiSumXY=0, HiN, HiMX, HiVX, HiMY, HiVY, HiR,
Wt, X, Exp2Z;
/* Lots of cryptic variable names! Here's the key:
Lo = outcome 2, Hi = outcome 3
X = attribute value, Y = input value (i.e., target residual)
M = mean, V = variance
N = weighted number of cases, K = count
R = correlation coefficient */
/* Special case when very few values */
if ( (Cases = Lp - Fp + 1) < 2 * MINSPLIT )
{
Verbosity(2, fprintf(Of, "Att %s\tinsufficient cases\n", AttName[Att]) )
return;
}
GEnv.BrSum[1] = GEnv.BrSumSq[1] = GEnv.BrFreq[1] =
GEnv.BrSum[2] = GEnv.BrSumSq[2] = GEnv.BrFreq[2] =
GEnv.BrSum[3] = GEnv.BrSumSq[3] = GEnv.BrFreq[3] = 0;
/* Isolate N/A values and sort the rest on the attribute value.
All cases are initially assigned to the high branch (outcome 3) */
Xp = Lp+1;
for ( i = Lp ; i >= Fp ; i-- )
{
Val = Resid(Case[i]);
Wt = CWeight(Case[i]);
if ( NotApplic(Case[i], Att) )
{
GEnv.BrSum[1] += Wt * Val;
GEnv.BrSumSq[1] += Wt * Val * Val;
GEnv.BrFreq[1] += Wt;
}
else
{
X = CVal(Case[i], Att);
Xp--;
SRec[Xp].V = X;
SRec[Xp].T = Val;
SRec[Xp].W = Wt;
HiSumX += Wt * X;
HiSumXX += Wt * X * X;
HiSumY += Wt * Val;
HiSumYY += Wt * Val * Val;
HiSumXY += Wt * X * Val;
GEnv.BrFreq[3] += Wt;
}
}
Cachesort(Xp, Lp);
/* Try possible cuts between items i and i+1 and determine the
gain of the split in each case */
Edge = (Cases >= 3 * MINITEMS ? MINITEMS : MINSPLIT );
ForEach(i, Xp, Lp - Edge)
{
X = SRec[i].V;
Val = SRec[i].T;
Wt = SRec[i].W;
GEnv.BrFreq[2] += Wt;
GEnv.BrFreq[3] -= Wt;
LoSumX += Wt * X;
LoSumXX += Wt * X * X;
LoSumY += Wt * Val;
LoSumYY += Wt * Val * Val;
LoSumXY += Wt * X * Val;
HiSumX -= Wt * X;
HiSumXX -= Wt * X * X;
HiSumY -= Wt * Val;
HiSumYY -= Wt * Val * Val;
HiSumXY -= Wt * X * Val;
if ( SRec[i+1].V > SRec[i].V && i >= Xp + Edge - 1 )
{
/* Possible cut here */
LoK = i - Xp + 1;
LoN = GEnv.BrFreq[2];
LoMX = LoSumX / LoN;
LoVX = LoSumXX / LoN - LoMX * LoMX;
LoMY = LoSumY / LoN;
LoVY = LoSumYY / LoN - LoMY * LoMY;
LoR = (LoSumXY - LoSumX * LoSumY / LoN) /
(LoN * sqrt(LoVX * LoVY + 1E-10));
Exp2Z = ( LoK < 6 ? 1E38 : exp(2 * 1.96 * sqrt(1.0 / (LoK - 3))) );
if ( fabs(LoR) < (Exp2Z - 1) / (Exp2Z + 1) ) LoR = 0;
HiK = Lp - i;
HiN = GEnv.BrFreq[3];
HiMX = HiSumX / HiN;
HiVX = HiSumXX / HiN - HiMX * HiMX;
HiMY = HiSumY / HiN;
HiVY = HiSumYY / HiN - HiMY * HiMY;
HiR = (HiSumXY - HiSumX * HiSumY / HiN) /
(HiN * sqrt(HiVX * HiVY + 1E-10));
Exp2Z = ( HiK < 6 ? 1E38 : exp(2 * 1.96 * sqrt(1.0 / (HiK - 3))) );
if ( fabs(HiR) < (Exp2Z - 1) / (Exp2Z + 1) ) HiR = 0;
/* Record the sums of squares of the imagined residuals */
GEnv.BrSumSq[2] = (1 - LoR * LoR) * LoN * LoVY;
GEnv.BrSumSq[3] = (1 - HiR * HiR) * HiN * HiVY;
if ( (ThisGain = ComputeGain(Node)) > BestGain )
{
BestGain = ThisGain;
BestI = i;
}
}
}
/* If there is a gain, set the best cut */
if ( BestGain > 0 )
{
GEnv.Gain[Att] = BestGain;
LowVal = SRec[BestI].V;
HighVal = SRec[BestI+1].V;
/* Set threshold, making sure that rounding problems do not
cause it to reach upper value */
if ( (GEnv.Bar[Att] = (ContValue) (0.5 * (LowVal + HighVal))) >= HighVal )
{
GEnv.Bar[Att] = LowVal;
}
Verbosity(2,
fprintf(Of, "Att %s\tcut=%.3f, gain %.3f\n",
AttName[Att], GEnv.Bar[Att], GEnv.Gain[Att]))
}
else
{
GEnv.Gain[Att] = None;
Verbosity(2, fprintf(Of, "Att %s\tno gain\n", AttName[Att]))
}
}
/*************************************************************************/
/* */
/* Change a leaf into a test on a continuous attribute */
/* */
/*************************************************************************/
void ContinTest(Tree Node, Attribute Att, float Cut)
/* ---------- */
{
Sprout(Node, 3);
Node->NodeType = BrThresh;
Node->Tested = Att;
Node->Cut = Cut;
}
/*************************************************************************/
/* */
/* Adjust thresholds of all continuous attributes so that cut */
/* values appear in the data */
/* */
/*************************************************************************/
Boolean Sorted;
void AdjustAllThresholds(Tree T)
/* ------------------- */
{
Attribute Att;
ForEach(Att, 1, MaxAtt)
{
if ( Continuous(Att) )
{
Sorted = false;
AdjustThresholds(T, Att);
}
}
}
void AdjustThresholds(Tree T, Attribute Att)
/* ---------------- */
{
DiscrValue v;
CaseNo i;
if ( T->NodeType == BrThresh && T->Tested == Att )
{
if ( ! Sorted )
{
/* Sort the attribute values. Leaving N/A values doesn't
matter since N/A will not lie between two real values */
ForEach(i, 0, MaxCase)
{
SRec[i].V = CVal(Case[i], Att);
}
Cachesort(0, MaxCase);
Sorted = true;
}
T->Cut = GreatestValueBelow(T->Cut);
}
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
AdjustThresholds(T->Branch[v], Att);
}
}
}
/*************************************************************************/
/* */
/* Return the greatest value of attribute Att below threshold Th. */
/* (Assumes values of Att have been sorted in SRec.) */
/* */
/*************************************************************************/
ContValue GreatestValueBelow(ContValue Th)
/* ------------------ */
{
CaseNo Low, High, Mid;
Low = 0;
High = MaxCase;
while ( Low < High )
{
Mid = (Low + High + 1) / 2;
if ( SRec[Mid].V > Th )
{
High = Mid - 1;
}
else
{
Low = Mid;
}
}
return SRec[Low].V;
}
cubist.c 0000664 0000764 0000764 00000014400 11501534415 012423 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Main routine, Cubist */
/* -------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#include
#include
#define SetOpt(V) V = strtod(OptArg, &EndPtr);\
if ( ! EndPtr || *EndPtr != '\00' ) break;\
ArgOK = true
#define SetIOpt(V) V = strtol(OptArg, &EndPtr, 10);\
if ( ! EndPtr || *EndPtr != '\00' ) break;\
ArgOK = true
int main(int Argc, char *Argv[])
/* ---- */
{
int o;
extern String OptArg, Option;
char *EndPtr;
Boolean FirstTime=true, ArgOK;
double StartTime, SumCWt=0;
FILE *F;
CaseNo SaveMaxCase, i, NCWt=0;
Attribute Att;
struct rlimit RL;
/* Make sure there is a largish runtime stack */
getrlimit(RLIMIT_STACK, &RL);
RL.rlim_cur = Max(RL.rlim_cur, 20 * 1024 * 1024);
if ( RL.rlim_max > 0 ) /* -1 if unlimited */
{
RL.rlim_cur = Min(RL.rlim_max, RL.rlim_cur);
}
setrlimit(RLIMIT_STACK, &RL);
/* Check for output to be saved to a file */
if ( Argc > 2 && ! strcmp(Argv[Argc-2], "-o") )
{
Of = fopen(Argv[Argc-1], "w");
Argc -= 2;
}
if ( ! Of )
{
Of = stdout;
}
KRInit = time(0) & 07777;
StartTime = ExecTime();
PrintHeader("");
/* Process options */
while ( (o = ProcessOption(Argc, Argv, "f+v+r+e+ian+uS+I+X+C+h")) )
{
if ( FirstTime )
{
fprintf(Of, T_OptHeader);
FirstTime = false;
}
ArgOK = false;
switch (o)
{
case 'f': FileStem = OptArg;
fprintf(Of, T_OptApplication, FileStem);
ArgOK = true;
break;
#ifdef VerbOpt
case 'v':
VERBOSITY = atoi(OptArg);
fprintf(Of, "\tVerbosity level %d\n", VERBOSITY);
ArgOK = true;
break;
#endif
case 'r': SetIOpt(MAXRULES);
fprintf(Of, T_OptMaxRules, MAXRULES);
Check(MAXRULES, 1, 1000000);
break;
case 'e': SetOpt(EXTRAP);
fprintf(Of, T_OptExtrap, EXTRAP);
Check(EXTRAP, 0, 100);
EXTRAP /= 100;
break;
case 'i': USEINSTANCES = true;
fprintf(Of, T_OptUseInst);
ArgOK = true;
break;
case 'a': CHOOSEMODE = USEINSTANCES = true;
fprintf(Of, T_OptAllowInst);
ArgOK = true;
break;
case 'n': USEINSTANCES = true;
SetIOpt(NN);
fprintf(Of, TX_NNeighbors(NN));
Check(NN, 1, NNMAX);
ArgOK = true;
break;
case 'u': UNBIASED = true;
fprintf(Of, T_NoBias);
ArgOK = true;
break;
case 'S': SetOpt(SAMPLE);
fprintf(Of, T_OptSampling, SAMPLE);
Check(SAMPLE, 0.1, 99.9);
SAMPLE /= 100;
break;
case 'I': SetIOpt(KRInit);
fprintf(Of, T_OptSeed, KRInit);
KRInit = KRInit & 07777;
break;
case 'X': SetIOpt(FOLDS);
fprintf(Of, T_OptXval, FOLDS);
Check(FOLDS, 1, 1000);
XVAL = true;
break;
case 'C': SetIOpt(MEMBERS);
fprintf(Of, T_OptCttee, MEMBERS);
Check(MEMBERS, 1, 100);
break;
}
if ( ! ArgOK )
{
if ( o != 'h' )
{
fprintf(Of, T_UnregnizedOpt, Option,
( ! OptArg || OptArg == Option+2 ? "" : OptArg ));
fprintf(Of, T_SummaryOpts);
}
fprintf(Of, T_ListOpts);
Goodbye(1);
}
}
/* Get information on training data */
if ( ! (F = GetFile(".names", "r")) ) Error(NOFILE, "", "");
GetNames(F);
fprintf(Of, T_TargetAtt, AttName[ClassAtt]);
NotifyStage(READDATA);
Progress(-1.0);
if ( ! (F = GetFile(".data", "r")) ) Error(NOFILE, "", "");
GetData(F, true, false);
fprintf(Of, TX_ReadData(MaxCase+1, MaxAtt, FileStem));
if ( XVAL && (F = GetFile(".test", "r")) )
{
SaveMaxCase = MaxCase;
GetData(F, false, false);
fprintf(Of, TX_ReadTest(MaxCase-SaveMaxCase, FileStem));
}
/* Check whether case weight attribute appears. If it does,
normalize values to average 1, and replace N/A values and
values <= 0 with average value 1 */
if ( CWtAtt )
{
fprintf(Of, T_CWtAtt);
/* Find average case weight value */
ForEach(i, 0, MaxCase)
{
if ( ! NotApplic(Case[i], CWtAtt) && CVal(Case[i], CWtAtt) > 0 )
{
SumCWt += CVal(Case[i], CWtAtt);
NCWt += 1;
}
}
AvCWt = ( NCWt > 0 ? SumCWt / NCWt : 1 );
ForEach(i, 0, MaxCase)
{
if ( ! NotApplic(Case[i], CWtAtt) && CVal(Case[i], CWtAtt) > 0 )
{
CVal(Case[i], CWtAtt) /= AvCWt;
}
else
{
CVal(Case[i], CWtAtt) = 1;
}
}
}
else
{
AvCWt = 1;
}
/* Note any attribute exclusions/inclusions */
if ( AttExIn )
{
fprintf(Of, "%s", ( AttExIn == -1 ? T_AttributesOut : T_AttributesIn ));
ForEach(Att, 1, MaxAtt)
{
if ( ( StatBit(Att, SKIP) > 0 ) == ( AttExIn == -1 ) )
{
fprintf(Of, " %s\n", AttName[Att]);
}
}
}
/* Build and evaluate cubist model */
InitialiseEnvData();
if ( XVAL )
{
CrossVal();
}
else
{
SingleCttee();
}
#ifdef VerbOpt
Cleanup();
#endif
fprintf(Of, T_Time, ExecTime() - StartTime);
return 0;
}
defns.i 0000664 0000764 0000764 00000053455 11501534415 012254 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Definitions used in Cubist */
/* -------------------------- */
/* */
/*************************************************************************/
#define CUBIST
#define RELEASE "2.07 GPL Edition"
#include
#include
#include
#include
#include
#include
#include
#include
#include "text.i"
/*************************************************************************/
/* */
/* Definitions dependent on cc options */
/* */
/*************************************************************************/
#define Goodbye(x) {Cleanup(); exit(x);}
#ifdef VerbOpt
#include
#define Verbosity(d,s) if(VERBOSITY >= d) {s;}
#define Free(x) {free(x); x=0;}
#else
#define assert(x)
#define Verbosity(d,s)
#define Free(x) free(x)
#endif
/*************************************************************************/
/* */
/* Constants, macros etc. */
/* */
/*************************************************************************/
#define MINSPLIT 3 /* min branch size for initial tree */
#define MINFRACT 0.001 /* min fraction of cases covered by rule */
#define MAXN 20 /* max neighbors allowing for ties */
#define NNMAX 9 /* upper limit for NN */
#define EVALSAMPLE 10000 /* sampling training data for evaluation */
#define Nil 0 /* null pointer */
#define false 0
#define true 1
#define None -1
#define Epsilon 1E-4
#define Width 80 /* approx max width of output */
#define EXCLUDE 1 /* special attribute status: do not use */
#define SKIP 2 /* do not use in models */
#define DISCRETE 4 /* ditto: collect values as data read */
#define ORDERED 8 /* ditto: ordered discrete values */
#define DATEVAL 16 /* ditto: YYYY/MM/DD or YYYY-MM-DD */
#define STIMEVAL 32 /* ditto: HH:MM:SS */
#define TSTMPVAL 64 /* date time */
#define BrDiscr 1
#define BrThresh 2
#define BrSubset 3
#define Plural(n) ((n) != 1 ? "s" : "")
//#define Alloc(N,T) (T *) Pmalloc((N)*sizeof(T))
#define AllocZero(N,T) (T *) Pcalloc(N, sizeof(T))
#define Alloc(N,T) AllocZero(N,T) /* for safety */
#define Realloc(V,N,T) V = (T *) Prealloc(V, (N)*sizeof(T))
#define Max(a,b) ((a)>(b) ? a : b)
#define Min(a,b) ((a)<(b) ? a : b)
#define Log2 0.69314718055994530942
#define Log(x) ((x) <= 0 ? 0.0 : log((double)x) / Log2)
#define Bit(b) (1 << (b))
#define In(b,s) ((s[(b) >> 3]) & Bit((b) & 07))
#define ClearBits(n,s) memset(s,0,n)
#define CopyBits(n,f,t) memcpy(t,f,n)
#define SetBit(b,s) (s[(b) >> 3] |= Bit((b) & 07))
#define ForEach(v,f,l) for(v=f ; v<=l ; ++v)
#define StatBit(a,b) (SpecialStatus[a]&(b))
#define Exclude(a) StatBit(a,EXCLUDE)
#define Skip(a) StatBit(a,EXCLUDE|SKIP)
#define Discrete(a) (MaxAttVal[a] || StatBit(a,DISCRETE))
#define Continuous(a) (! MaxAttVal[a] && ! StatBit(a,DISCRETE))
#define Ordered(a) StatBit(a,ORDERED)
#define DateVal(a) StatBit(a,DATEVAL)
#define TimeVal(a) StatBit(a,STIMEVAL)
#define TStampVal(a) StatBit(a,TSTMPVAL)
#define UNKNOWN 1.5777218104420236e-30 /* unlikely value! */
#define NA 1
#define NotApplic(c,a) (DVal(c,a)==NA)
#define NotApplicVal(AV) (AV._discr_val==NA)
#define FreeUnlessNil(p) if((p)!=Nil) Free(p)
#define CheckClose(f) if(f) {fclose(f); f=Nil;}
#define Space(s) (s==' ' || s=='\n' || s=='\r' || s=='\t')
#define SkipComment while ( (c = InChar(f)) != '\n' && c != EOF )
#define P1(x) (rint((x)*10) / 10)
#define Before(n1,n2) (n1->Tested < n2->Tested ||\
n1->Tested == n2->Tested && n1->Cut < n2->Cut)
#define Swap(a,b) {xab = Case[a]; Case[a] = Case[b]; Case[b] = xab;}
#define MinAttCoeff(a) (0.01 * GlobalSD / AttSD[a]) /* otherwise zero */
#define NOFILE 0
#define BADATTNAME 1
#define EOFINATT 2
#define SINGLEATTVAL 3
#define BADATTVAL 4
#define BADNUMBER 5
#define DUPATTNAME 6
#define NOMEM 8
#define TOOMANYVALS 9
#define BADDISCRETE 10
#define NOTARGET 11
#define BADTARGET 12
#define LONGNAME 13
#define HITEOF 14
#define MISSNAME 15
#define BADDATE 16
#define BADTIME 17
#define BADTSTMP 18
#define UNKNOWNATT 19
#define BADDEF1 20
#define BADDEF2 21
#define BADDEF3 22
#define SAMEATT 23
#define BADDEF4 24
#define MODELFILE 30
#define CWTATTERR 31
#define READDATA 1
#define GROUPDATA 2
#define ADDMODELS 3
#define SIMPLIFYGROUPS 4
#define FORMRULES 5
#define INDEXINSTANCES 6
#define SETNEIGHBORS 7
#define ASSESSCOMPOSITE 8
#define EVALTRAIN 9
#define READTEST 10
#define EVALTEST 11
#define CLEANUP 12
#define RESULTS 13
/*************************************************************************/
/* */
/* Type definitions */
/* */
/*************************************************************************/
typedef unsigned char Boolean, BranchType, *Set;
typedef char *String;
typedef int CaseNo, /* data item number */
CaseCount, /* count of cases */
DiscrValue, /* discrete attribute value (0 = ?) */
Attribute; /* attribute number, 1..MaxAtt */
#ifdef USEDOUBLE
typedef double ContValue; /* continuous attribute value */
#define PREC 14 /* precision */
#else
typedef float ContValue; /* continuous attribute value */
#define PREC 7 /* precision */
#endif
/* Attribute values are packed into a union:
DVal = (int) discrete value
CVal = (float) continuous value
SVal = (int) offset in IgnoredVals
Missing and non-applicable values are:
discrete:
not applicable: DVal = NA
missing: DVal = 0
continuous:
not applicable: DVal = NA
missing: CVal = UNKNOWN */
typedef union _attribute_value
{
ContValue _cont_val;
DiscrValue _discr_val;
}
AttValue, *DataRec;
#define CVal(Case,Attribute) Case[Attribute]._cont_val
#define DVal(Case,Attribute) Case[Attribute]._discr_val
#define XDVal(Case,Attribute) (Case[Attribute]._discr_val & 077777777)
#define SVal(Case,Attribute) Case[Attribute]._discr_val
#define Class(Case) (*Case)._cont_val
#define Resid(Case) Case[MaxAtt+1]._cont_val
#define PredSum(Case) Case[MaxAtt+1]._cont_val
#define PredCount(Case) Case[MaxAtt+2]._discr_val
#define PredVal(Case) Case[MaxAtt+1]._cont_val
#define DRef1(Case) Case[MaxAtt+1]._cont_val
#define DRef2(Case) Case[MaxAtt+2]._cont_val
#define CWeight(Case) ( CWtAtt ? CVal(Case,CWtAtt) : 1.0 )
typedef struct _env_rec
{
double *LocalModel, /* intermediate regression model */
*ValFreq, /* count of items with att value v */
BrFreq[4], /* for split */
*ValSum, /* sum class values for value v */
*ValSumSq, /* ditto sum squares */
BrSum[4], /* sum class values for branch b */
BrSumSq[4]; /* ditto sum squares */
Boolean *DoNotUse; /* atts precxluded from model */
float *Gain; /* gain from splitting on att */
ContValue *Bar; /* best threshold for contin att */
Boolean *Left; /* true if v is in left subset */
Set **Subset; /* subset s for att a */
Attribute *ModelAtt; /* atts used in current model */
int NModelAtt; /* number ditto */
double **xTx, /* [Att][Att] */
*xTy, /* [Att] */
**A, /* copy of xTx destroyed by inversion */
*B, /* copy of xTy ditto */
*BestModel, /* for SimplifyModel */
*Resid, /* [Case] */
*PResid, /* [Case] */
*Mean, /* [Att] */
*Var, /* [Att] */
*AvDev; /* [Att] */
Boolean *ZeroCoeff, /* true if coeff to be set to zero */
*SaveZero; /* for SimplifyModel */
DataRec *Filtered; /* items minus outliers */
}
EnvRec, *Env;
typedef struct _sort_rec
{
ContValue V, /* attribute value */
T, /* target value or residual */
W; /* weight */
}
SortRec;
typedef struct _treerec *Tree;
typedef struct _treerec
{
BranchType NodeType; /* 0 | BrDiscr | BrThresh | BrSubset */
CaseCount Cases; /* no of cases at this node */
double Mean, /* average of cases at this node */
SD, /* standard dev ditto */
*Model, /* model at this node */
*MCopy; /* copy used during smoothing */
Attribute Tested; /* attribute referenced in test */
int Forks; /* number of branches at this node */
ContValue Cut; /* threshold for continuous attribute */
Set *Subset; /* subsets of discrete values */
Tree *Branch; /* Branch[x] = subtree for outcome x */
float Params, /* parameters in this tree */
Coeffs, /* coefficients of leaf model */
TreeErr, /* error as tree */
LeafErr, /* error if replace by leaf */
Utility;
}
TreeRec;
#define Parent(T) T->Branch[0]
typedef int RuleNo; /* rule number */
typedef struct _condrec
{
BranchType NodeType; /* test type (see tree nodes) */
Attribute Tested; /* attribute tested */
ContValue Cut; /* threshold (if relevant) */
Set Subset; /* subset (if relevant) */
int TestValue; /* specified outcome of test */
}
CondRec, *Condition;
typedef struct _rulerec
{
RuleNo RNo; /* rule number */
int MNo, /* member number for committee models */
Size; /* number of conditions */
Condition *Lhs; /* conditions themselves */
double *Rhs; /* model given by rule */
CaseNo Cover; /* number of cases covered */
float Mean, /* mean value of cases matching rule */
LoVal, /* lowest value in data */
HiVal, /* highest value in data */
LoLim, /* lower bound on predictions */
HiLim, /* upper bound on predictions */
EstErr; /* estimated error */
}
RuleRec, *CRule;
typedef struct _oldrulerec
{
RuleNo RNo; /* rule number */
int Size; /* number of conditions */
Condition *Lhs; /* conditions themselves */
double *Rhs; /* model given by rule */
CaseNo Cover; /* number of cases covered */
float Mean, /* mean value of cases matching rule */
LoVal, /* lowest value in data */
HiVal, /* highest value in data */
LoLim, /* lower bound on predictions */
HiLim, /* upper bound on predictions */
EstErr; /* estimated error */
}
OldRuleRec;
typedef struct _rulesetrec
{
RuleNo SNRules; /* number of rules */
CRule *SRule; /* rules */
}
RuleSetRec, *RRuleSet;
typedef struct _indexrec *Index;
typedef struct _indexrec
{
Attribute Tested; /* split attribute for KD-tree */
ContValue Cut, /* threshold for continuous atts */
MinDRef[2], /* min reference distances */
MaxDRef[2]; /* max ditto */
CaseNo IFp, ILp; /* first and last item at leaf */
Index *SubIndex; /* subtrees */
}
IndexRec;
typedef struct _nnrec
{
int BestI[MAXN]; /* numbers of best instances */
float BestD[MAXN], /* distances to best instances */
*WorstBest, /* points to worst BestD */
*AttMinD; /* min attribute distance from case */
}
NNEnvRec;
typedef union _def_val
{
String _s_val; /* att val for comparison */
ContValue _n_val; /* number for arith */
}
DefVal;
typedef struct _def_elt
{
short _op_code; /* type of element */
DefVal _operand; /* string or numeric value */
}
DefElt, *Definition;
typedef struct _elt_rec
{
int Fi, /* index of first char of element */
Li; /* last ditto */
char Type; /* 'B', 'S', or 'N' */
}
EltRec;
#define DefOp(DE) DE._op_code
#define DefSVal(DE) DE._operand._s_val
#define DefNVal(DE) DE._operand._n_val
#define OP_ATT 0 /* opcodes */
#define OP_NUM 1
#define OP_STR 2
#define OP_MISS 3
#define OP_AND 10
#define OP_OR 11
#define OP_EQ 20
#define OP_NE 21
#define OP_GT 22
#define OP_GE 23
#define OP_LT 24
#define OP_LE 25
#define OP_SEQ 26
#define OP_SNE 27
#define OP_PLUS 30
#define OP_MINUS 31
#define OP_UMINUS 32
#define OP_MULT 33
#define OP_DIV 34
#define OP_MOD 35
#define OP_POW 36
#define OP_SIN 40
#define OP_COS 41
#define OP_TAN 42
#define OP_LOG 43
#define OP_EXP 44
#define OP_INT 45
#define OP_END 99
/*************************************************************************/
/* */
/* Function prototypes */
/* */
/*************************************************************************/
/* cubist.c */
int main(int, char *[]);
/* construct.c */
void SingleCttee(void);
void ConstructCttee(void);
RRuleSet ConstructRuleSet(int ModelNo);
void EvaluateCttee(RRuleSet *RS, Boolean Details);
void SampleTrainingCases(void);
void AttributeUsage(void);
void UpdateUsage(CRule R);
void NoteUsed(Attribute Att);
/* getnames.c */
Boolean ReadName(FILE *f, String s, int n, char ColonOpt);
void GetNames(FILE *Nf);
void ExplicitAtt(FILE *Nf);
int Which(String Val, String *List, int First, int Last);
void ListAttsUsed(void);
void FreeNamesData(void);
int InChar(FILE *f);
/* implicitatt.c */
void ImplicitAtt(FILE *Nf);
void ReadDefinition(FILE *f);
void Append(char c);
Boolean Expression(void);
Boolean Conjunct(void);
Boolean SExpression(void);
Boolean AExpression(void);
Boolean Term(void);
Boolean Factor(void);
Boolean Primary(void);
Boolean Atom(void);
Boolean Find(String S);
int FindOne(String *Alt);
Attribute FindAttName(void);
void DefSyntaxError(String Msg);
void DefSemanticsError(int Fi, String Msg, int OpCode);
void Dump(char OpCode, ContValue F, String S, int Fi);
void DumpOp(char OpCode, int Fi);
Boolean UpdateTStack(char OpCode, ContValue F, String S, int Fi);
AttValue EvaluateDef(Definition D, DataRec Case);
/* getdata.c */
void GetData(FILE *Df, Boolean Train, Boolean AllowUnknownTarget);
Boolean ReplaceUnknowns(DataRec Case, Boolean *AttMsg);
DataRec GetDataRec(FILE *Df, Boolean Train);
CaseNo CountData(FILE *Df);
int StoreIVal(String S);
void FreeData(DataRec *Case);
void CheckValue(DataRec Case, Attribute Att);
void FindLimits(Attribute Att, ContValue *Min, ContValue *Max);
/* predict.c */
float PredictValue(RRuleSet *Cttee, DataRec CaseDesc);
float RuleSetPrediction(RRuleSet RS, DataRec CaseDesc);
Boolean Matches(CRule R, DataRec Case);
float LinModel(double *Model, DataRec Case);
float RawLinModel(double *Model, DataRec Case);
void FindPredictedValues(RRuleSet *RS, CaseNo Fp, CaseNo Lp);
/* formtree.c */
void InitialiseEnvData(void);
void FreeEnvData(void);
void FindGlobalProperties(void);
void FormTree(CaseNo, CaseNo, int, Tree *, Tree);
void AddModels(CaseNo Fp, CaseNo Lp, Tree T, Tree Parent);
CaseNo Group(DiscrValue, CaseNo, CaseNo, Tree);
void Divide(Tree Node, CaseNo Fp, CaseNo Lp, int Level);
void AddSplitAtts(Tree T);
void AddDefAtts(void);
void FindModelAtts(double *Model);
/* discr.c */
void EvalDiscreteAtt(Tree, Attribute, CaseNo Fp, CaseNo Lp);
void EvalBinarySplit(Tree, Attribute, CaseNo Fp, CaseNo Lp);
void EvalSubsetSplit(Tree, Attribute, CaseNo Fp, CaseNo Lp);
void DiscreteTest(Tree Node, Attribute Att, Set *Subset);
/* contin.c */
void EvalContinuousAtt(Tree, Attribute, CaseNo Fp, CaseNo Lp);
void ContinTest(Tree Node, Attribute Att, float Cut);
void AdjustAllThresholds(Tree T);
void AdjustThresholds(Tree T, Attribute Att);
ContValue GreatestValueBelow(ContValue Th);
/* sort.c */
void Cachesort(CaseNo Fp, CaseNo Lp);
/* trees.c */
void FindDepth(Tree T);
void PrintTree(Tree T, String Title);
void Show(Tree T, int Sh);
void ShowBranch(int Sh, Tree T, DiscrValue v, DiscrValue BrNo);
int MaxLine(Tree SubTree);
void Indent(int Sh, int BrNo);
void FreeTree(Tree T);
Tree Leaf(CaseCount Cases, double Mean, double SD);
void Sprout(Tree T, DiscrValue Branches);
int TreeSize(Tree T);
int TreeLeaves(Tree T);
/* prunetree.c */
void Prune(Tree T);
void SetProperties(Tree T, Tree Parent);
void UnsproutAndUpdate(Tree Pruned, double ExtraErr, double ExtraParams);
void Unsprout(Tree T);
void SmoothModels(Tree T, Tree Parent, CaseNo Fp, CaseNo Lp);
double ErrVariance(double *Model, CaseNo Fp, CaseNo Lp, double *Err);
double Smooth(double PVal, int PCases, double LVal, int LCases);
void FindErrors(Tree T, CaseNo Fp, CaseNo Lp);
void CValToStr(ContValue CV, Attribute Att, String DS);
Boolean FindWeakestSubtree(Tree T);
float TreeValue(Tree T, DataRec Case);
float MedianResid(CaseNo Fp, CaseNo Lp, double Want);
/* regress.c */
void Regress(CaseNo Fp, CaseNo Lp, double *Model);
void BuildTables(CaseNo Fp, CaseNo Lp);
void Solve(double *Model);
void FindActiveAtts(void);
void AddRow(double *Model, short From, short To, double Factor);
void ExchangeRow(double *Model, short From, short To);
int CountCoeffs(double *Model);
void SimplifyModel(DataRec *D, CaseNo Fp, CaseNo Lp, double *Model);
/* stats.c */
double AverageDev(float Mean, CaseNo Fp, CaseNo Lp);
double SD(double N, double Sum, double SumSq);
double ComputeGain(Tree Node);
double AverageErr(DataRec *D, CaseNo Fp, CaseNo Lp, double *Model);
double EstimateErr(double Val, double NData, float NParam);
/* utility.c */
void PrintHeader(String Title);
char ProcessOption(int Argc, char **Argv, char *Str);
void *Pmalloc(size_t Bytes);
void *Prealloc(void *Present, size_t Bytes);
void *Pcalloc(size_t Number, unsigned int Size);
DataRec NewCase(void);
void FreeCases(void);
void FreeLastCase(DataRec Case);
void FreeVector(void **V, int First, int Last);
double KRandom(void);
void ResetKR(int KRInit);
void Error(int ErrNo, String S1, String S2);
String CaseLabel(CaseNo N);
FILE * GetFile(String Extension, String RW);
double ExecTime(void);
int Denominator(ContValue Val);
int FracBase(Attribute Att);
int GetInt(String S, int N);
int DateToDay(String DS);
void DayToDate(int DI, String Date);
int TimeToSecs(String TS);
void SecsToTime(int Secs, String Time);
void SetTSBase(int y);
int TStampToMins(String TS);
void Check(float Val, float Low, float High);
void Cleanup(void);
#ifdef UTF8
int UTF8CharWidth(unsigned char *U);
int wcwidth(wchar_t ucs);
int wcswidth(const wchar_t *pwcs, size_t n);
#endif
/* formrules.c */
RRuleSet FormRules(Tree T);
void TreeParameters(Tree T, int D);
void Scan(Tree T);
void PushCondition(void);
void PopCondition(void);
void PruneRule(Condition Cond[], float InitCoeffs);
void UpdateCount(int d, CaseNo i, double *Total, double *PredErr);
void ProcessLists(void);
void AddToList(CaseNo *List, CaseNo N);
void DeleteFromList(CaseNo *Before, CaseNo N);
int SingleFail(CaseNo i);
void RemoveBias(CRule R, int Coeffs);
void OrderRules(void);
void FreeFormRuleData(void);
/* rules.c */
Boolean NewRule(Condition Cond[], int NConds, Boolean *Deleted,
CaseNo Cover, float Mean, float LoVal, float HiVal,
float EstErr, double *Model);
Boolean SameRule(RuleNo r, Condition Cond[], int NConds);
void ReleaseRule(CRule R);
void FreeCttee(RRuleSet *Cttee);
void PrintRules(RRuleSet, String);
void PrintRule(CRule R);
void PrintCondition(Condition C);
Boolean Satisfies(DataRec CaseDesc, Condition OneCond);
/* modelfiles.c */
void CheckFile(String Extension, Boolean Write);
void WriteFilePrefix(String Extension);
void ReadFilePrefix(String Extension);
void SaveDiscreteNames(void);
void SaveCommittee(RRuleSet *Cttee, String Extension);
void SaveRules(RRuleSet RS);
void AsciiOut(String Pre, String S);
void ReadHeader(void);
RRuleSet *GetCommittee(String Extension);
RRuleSet InRules(void);
CRule InRule(void);
Condition InCondition(void);
int ReadProp(char *Delim);
String RemoveQuotes(String S);
Set MakeSubset(Attribute Att);
/* update.c (Unix) or frontend.c (WIN32) */
void NotifyStage(int);
void Progress(float);
/* instance.c */
void InitialiseInstances(RRuleSet *Cttee);
void SetParameters(RRuleSet *Cttee);
void CheckForms(RRuleSet *Cttee);
void CopyInstances(void);
float NNEstimate(RRuleSet *Cttee, DataRec Case);
float Distance(DataRec Case1, DataRec Case2, float Thresh);
void CheckDistance(DataRec Case, CaseNo Saved);
void FindNearestNeighbors(DataRec Case);
float AverageNeighbors(RRuleSet *Cttee, DataRec Case);
Index BuildIndex(CaseNo Fp, CaseNo Lp);
void ScanIndex(DataRec Case, Index Node, float MinD);
void SwapInstance(CaseNo A, CaseNo B);
void FreeIndex(Index Node);
void FreeInstances(void);
/* xval.c */
void CrossVal(void);
void Prepare(void);
void Shuffle(int *);
void Summary(void);
discr.c 0000664 0000764 0000764 00000014604 11501534415 012244 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Evaluation of a test on a discrete valued attribute */
/* --------------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define Reasonable ((GEnv.BrFreq[1]>=MINSPLIT) + (GEnv.BrFreq[2]>=MINSPLIT) +\
(GEnv.BrFreq[3]>=MINSPLIT) >= 2)
/*************************************************************************/
/* */
/* Set Gain[] for discrete partition of items Fp to Lp */
/* */
/*************************************************************************/
void EvalDiscreteAtt(Tree Node, Attribute Att, CaseNo Fp, CaseNo Lp)
/* --------------- */
{
if ( MaxAttVal[Att] == 3 )
{
EvalBinarySplit(Node, Att, Fp, Lp);
}
else
{
EvalSubsetSplit(Node, Att, Fp, Lp);
}
Verbosity(2,
if ( GEnv.Gain[Att] >= 0 )
{
fprintf(Of, "Att %s: gain %.3f\n", AttName[Att], GEnv.Gain[Att]);
}
else
{
fprintf(Of, "Att %s: no gain\n", AttName[Att]);
})
}
/*************************************************************************/
/* */
/* Special case of binary split */
/* */
/*************************************************************************/
void EvalBinarySplit(Tree Node, Attribute Att, CaseNo Fp, CaseNo Lp)
/* --------------- */
{
CaseNo i;
DiscrValue v;
ContValue Cv;
double Wt;
ForEach(v, 1, 3)
{
GEnv.BrFreq[v] = GEnv.BrSum[v] = GEnv.BrSumSq[v] = 0;
}
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
v = DVal(Case[i], Att);
GEnv.BrFreq[v] += Wt;
GEnv.BrSum[v] += Wt * (Cv = Resid(Case[i]));
GEnv.BrSumSq[v] += Wt * Cv * Cv;
}
GEnv.Gain[Att] = ( Reasonable ? ComputeGain(Node) : None );
}
/*************************************************************************/
/* */
/* Divide attribute values into two subsets */
/* */
/*************************************************************************/
void EvalSubsetSplit(Tree Node, Attribute Att, CaseNo Fp, CaseNo Lp)
/* --------------- */
{
CaseNo i;
DiscrValue v, sv, Cycle;
ContValue Cv;
int Bytes;
double Wt, ThisGain;
ForEach(v, 1, MaxAttVal[Att])
{
GEnv.ValFreq[v] = GEnv.ValSum[v] = GEnv.ValSumSq[v] = 0;
}
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
v = DVal(Case[i], Att);
GEnv.ValFreq[v] += Wt;
GEnv.ValSum[v] += Wt * (Cv = Resid(Case[i]));
GEnv.ValSumSq[v] += Wt * Cv * Cv;
}
/* Initialise branch counts and sums */
GEnv.BrFreq[1] = GEnv.ValFreq[1];
GEnv.BrSum[1] = GEnv.ValSum[1];
GEnv.BrSumSq[1] = GEnv.ValSumSq[1];
Bytes = (MaxAttVal[Att]>>3) + 1;
ClearBits(Bytes, GEnv.Subset[Att][1]);
SetBit(1, GEnv.Subset[Att][1]);
ForEach(v, 2, 3)
{
GEnv.BrFreq[v] = GEnv.BrSum[v] = GEnv.BrSumSq[v] = 0;
}
/* Examine subsets, starting with all values in the left branch.
At each iteration, move the value with the highest mean from
the left branch to the right branch and check the gain.
(In the case of ordered attributes, the value moved is the
rightmost value in the left branch.) Save the best gain so far
in GEnv.Subset[Att]. */
ForEach(v, 2, MaxAttVal[Att])
{
GEnv.BrFreq[2] += GEnv.ValFreq[v];
GEnv.BrSum[2] += GEnv.ValSum[v];
GEnv.BrSumSq[2] += GEnv.ValSumSq[v];
}
ForEach(v, 2, MaxAttVal[Att])
{
GEnv.Left[v] = ( GEnv.ValFreq[v] > 0 );
}
ForEach(Cycle, 2, MaxAttVal[Att])
{
if ( Ordered(Att) )
{
for ( sv = MaxAttVal[Att] ; sv > 1 && ! GEnv.Left[sv] ; sv-- )
;
}
else
{
sv = 0;
ForEach(v, 2, MaxAttVal[Att])
{
if ( GEnv.Left[v] &&
( ! sv ||
GEnv.ValSum[v] / GEnv.ValFreq[v] >
GEnv.ValSum[sv] / GEnv.ValFreq[sv] ) )
{
sv = v ;
}
}
}
if ( sv < 2 ) break;
GEnv.Left[sv] = false;
GEnv.BrFreq[2] -= GEnv.ValFreq[sv];
GEnv.BrSum[2] -= GEnv.ValSum[sv];
GEnv.BrSumSq[2] -= GEnv.ValSumSq[sv];
GEnv.BrFreq[3] += GEnv.ValFreq[sv];
GEnv.BrSum[3] += GEnv.ValSum[sv];
GEnv.BrSumSq[3] += GEnv.ValSumSq[sv];
ThisGain = ( Reasonable ? ComputeGain(Node) : None );
if ( ThisGain > GEnv.Gain[Att] )
{
GEnv.Gain[Att] = ThisGain;
/* Record in GEnv.Subset[Att] */
ClearBits(Bytes, GEnv.Subset[Att][2]);
ClearBits(Bytes, GEnv.Subset[Att][3]);
ForEach(v, 2, MaxAttVal[Att])
{
if ( GEnv.ValFreq[v] > 0 )
{
SetBit(v, GEnv.Subset[Att][ ( GEnv.Left[v] ? 2 : 3 ) ]);
}
}
}
}
}
/*************************************************************************/
/* */
/* Construct and return a node for a test on a discrete attribute */
/* */
/*************************************************************************/
void DiscreteTest(Tree Node, Attribute Att, Set *Subset)
/* ------------ */
{
int S, Bytes;
Sprout(Node, 3);
Node->Tested = Att;
if ( MaxAttVal[Att] == 3 )
{
Node->NodeType = BrDiscr;
}
else
{
Node->NodeType = BrSubset;
Bytes = (MaxAttVal[Att]>>3) + 1;
Node->Subset = AllocZero(4, Set);
ForEach(S, 1, 3)
{
Node->Subset[S] = Alloc(Bytes, unsigned char);
CopyBits(Bytes, Subset[S], Node->Subset[S]);
}
}
}
extern.i 0000664 0000764 0000764 00000005415 11501534415 012453 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Data defined in global.c */
/* ------------------------ */
/* */
/*************************************************************************/
extern Attribute
ClassAtt,
LabelAtt,
CWtAtt;
extern char *IgnoredVals;
extern int IValsSize,
IValsOffset;
extern int MaxAtt,
MaxDiscrVal,
Precision,
MaxLabel,
LineNo,
ErrMsgs,
AttExIn,
TSBase;
extern CaseNo MaxCase;
extern DataRec *Case;
extern DataRec *SaveCase,
*Blocked;
extern CaseNo SaveMaxCase;
extern DiscrValue
*MaxAttVal,
*Modal;
extern char *SpecialStatus;
extern Definition
*AttDef;
extern Attribute
**AttDefUses;
extern String *AttName,
**AttValName;
extern FILE *Of;
extern String FileStem;
extern ContValue
*AttMean,
*AttSD,
*AttMaxVal,
*AttMinVal,
*AttPref,
Ceiling,
Floor,
AvCWt;
extern float ErrReduction;
double *AttUnit;
extern int *AttPrec;
extern DataRec *Instance,
Ref[2];
extern CaseNo MaxInstance;
extern Index KDTree;
extern NNEnvRec
GNNEnv;
extern float *RSPredVal;
extern EnvRec GEnv;
extern Tree TempMT;
extern SortRec *SRec;
extern float GlobalMean,
GlobalSD,
GlobalErr;
extern char Fn[512];
extern FILE *Mf,
*Pf;
extern CRule *Rule;
extern RuleNo NRules;
extern int RuleSpace;
extern RRuleSet
*Cttee;
extern int VERBOSITY,
FOLDS,
NN,
MEMBERS;
extern float MAXD;
extern Boolean XVAL,
CHOOSEMODE,
USEINSTANCES,
UNBIASED;
extern float SAMPLE;
extern int KRInit;
extern Boolean LOCK;
extern CaseCount
MINITEMS;
extern int MAXRULES;
extern float EXTRAP;
formrules.c 0000664 0000764 0000764 00000036461 11501534415 013163 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Form a set of rules from a model tree */
/* ------------------------------------- */
/* */
/* The cases are partitionen into sublists: */
/* * Fail0: those cases that satisfy all undeleted conditions */
/* * Fail1: those that satisfy all but one of the above */
/* * FailMany: the remaining cases */
/* Lists are implemented via Succ; Succ[i] is the number of the */
/* case that follows case i. */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
double *Total=Nil, /* [Condition] */
*PredErr=Nil, /* [Condition] */
*Model;
Boolean **CondFailedBy=Nil, /* [Condition][CaseNo] */
*Deleted=Nil; /* [Condition] */
Condition *Stack=Nil;
int Leaves,
MaxDepth,
NCond,
Bestd;
short *NFail=Nil, /* NFail[i] = conditions failed by i */
*LocalNFail=Nil; /* copy used during rule pruning */
CaseNo *Succ=Nil, /* case following case i */
Fail0, /* first satisfying all conditions */
Fail1, /* ditto all but one */
FailMany; /* first case failing 2+ conditions */
float *CPredVal=Nil; /* raw model values for each case */
RRuleSet FormRules(Tree T)
/* --------- */
{
CaseNo i;
RRuleSet RS;
/* Record values predicted by tree */
ForEach(i, 0, MaxCase)
{
PredSum(Case[i]) = TreeValue(T, Case[i]);
PredCount(Case[i]) = 1;
}
/* Find essential parameters and allocate storage */
MaxDepth = Leaves = 0;
TreeParameters(T, 0);
Total = Alloc(MaxDepth+2, double);
PredErr = Alloc(MaxDepth+2, double);
CondFailedBy = AllocZero(MaxDepth+2, Boolean *);
Deleted = AllocZero(MaxDepth+2, Boolean);
Stack = AllocZero(MaxDepth+2, Condition);
ForEach(i, 0, MaxDepth+1)
{
CondFailedBy[i] = AllocZero(MaxCase+1, Boolean);
}
NFail = AllocZero(MaxCase+1, short);
LocalNFail = AllocZero(MaxCase+1, short);
Succ = Alloc(MaxCase+1, CaseNo);
NRules = RuleSpace = 0;
CPredVal = Alloc(MaxCase+1, float);
/* Extract and prune paths from root to leaves */
NCond = 0;
Scan(T);
OrderRules();
RS = Alloc(1, RuleSetRec);
RS->SNRules = NRules;
RS->SRule = Rule; Rule = Nil;
/* Deallocate storage */
FreeFormRuleData();
return RS;
}
/*************************************************************************/
/* */
/* Find the maximum depth and the number of leaves in tree T */
/* with initial depth D */
/* */
/*************************************************************************/
void TreeParameters(Tree T, int D)
/* -------------- */
{
DiscrValue v;
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
TreeParameters(T->Branch[v], D+1);
}
}
else
{
Leaves++;
if ( D > MaxDepth ) MaxDepth = D;
}
}
/*************************************************************************/
/* */
/* Extract paths from tree T at depth D, and process them */
/* */
/*************************************************************************/
void Scan(Tree T)
/* ---- */
{
DiscrValue v;
Condition Term;
if ( T->NodeType )
{
NCond++;
Term = AllocZero(1, CondRec);
Term->NodeType = T->NodeType;
Term->Tested = T->Tested;
Term->Cut = T->Cut;
Stack[NCond] = Term;
ForEach(v, 1, T->Forks)
{
Stack[NCond]->TestValue = v;
if ( Term->NodeType == BrSubset ) Term->Subset = T->Subset[v];
/* Adjust number of failed conditions */
PushCondition();
Scan(T->Branch[v]);
/* Reset number of failed conditions */
PopCondition();
}
/* Free local storage */
Free(Term);
NCond--;
}
else
if ( T->Cases >= 1 )
{
memcpy(LocalNFail, NFail, (MaxCase+1) * sizeof(short));
/* Prune the current path */
Model = T->Model;
PruneRule(Stack, T->Coeffs);
if ( ! T->NodeType ) Progress((float) T->Cases);
}
}
/*************************************************************************/
/* */
/* Update NFail when a condition is added to/removed from Stack */
/* */
/*************************************************************************/
void PushCondition()
/* ------------- */
{
int i;
ForEach(i, 0, MaxCase)
{
if ( (CondFailedBy[NCond][i] = ! Satisfies(Case[i], Stack[NCond])) )
{
NFail[i]++;
}
}
}
void PopCondition()
/* ------------- */
{
int i;
ForEach(i, 0, MaxCase)
{
if ( CondFailedBy[NCond][i] )
{
NFail[i]--;
}
}
}
/*************************************************************************/
/* */
/* Prune the rule given by the conditions Cond, and the number of */
/* conditions NCond, and add the resulting rule to the current */
/* ruleset if it is sufficiently accurate */
/* */
/*************************************************************************/
void PruneRule(Condition Cond[], float InitCoeffs)
/* --------- */
{
int d, id, Bestid, Remaining=NCond;
ContValue Val, LoVal=1E38, HiVal=-1E38, Wt;
double Sum=0, SumWt=0;
CaseNo i;
CaseCount Cover=0;
FindModelAtts(Model);
Bestd = 0;
ForEach(d, 0, NCond)
{
Deleted[d] = false;
}
/* Find conditions to delete */
Verbosity(1, fprintf(Of, "\n Pruning rule"))
while ( true )
{
/* Initialise or update fail lists, totals, and predicted errors */
ProcessLists();
if ( Remaining == 1 ) break;
/* See whether a condition can be deleted */
Bestd = id = 0;
Verbosity(1,
fprintf(Of, "\n Used Err \tAbsent condition\n"))
ForEach(d, 0, NCond)
{
if ( Deleted[d] ) continue;
Verbosity(1,
fprintf(Of, " %7g%7.1f",
rint(Total[d] * 10) / 10, PredErr[d]))
if ( ! d )
{
Verbosity(1, fprintf(Of, "\t\n"))
}
else
{
id++;
Verbosity(1, PrintCondition(Cond[d]))
/* Bestd identifies the condition whose removal would
lead to the greatest improvement in the newly-covered
cases */
if ( PredErr[d] >= 0 &&
( ! Bestd || PredErr[d] > PredErr[Bestd] ) )
{
Bestd = d;
Bestid = id;
}
}
}
if ( ! Bestd ) break;
Verbosity(1, fprintf(Of, "\teliminate test %d\n", Bestid))
Deleted[Bestd] = true;
Remaining--;
}
/* Check whether pruned rule is viable */
if ( NCond && ! Remaining ) return;
/* Find lowest and highest value among cases covered by this rule,
and the average local error */
for ( i = Fail0 ; i >= 0 ; )
{
Cover++;
Wt = CWeight(Case[i]);
SumWt += Wt;
Sum += Wt * (Val = Class(Case[i]));
LoVal = Min(LoVal, Val);
HiVal = Max(HiVal, Val);
i = Succ[i];
}
/* Add this as a possible new rule */
PredErr[0] = EstimateErr(PredErr[0]/Total[0], Cover, InitCoeffs);
if ( NewRule(Cond, NCond, Deleted, Cover, Sum / SumWt, LoVal, HiVal,
PredErr[0], Model) )
{
/* Adjust average predictions for new cases covered by this rule.
Do not adjust for cases covered by the initial rule, since the
value was entered at initialisation */
for ( i = Fail0 ; i >= 0 ; )
{
if ( NFail[i] ) /* not covered by initial rule */
{
PredSum(Case[i]) += ( CPredVal[i] < LoVal ? LoVal :
CPredVal[i] > HiVal ? HiVal :
CPredVal[i] );
PredCount(Case[i])++;
}
i = Succ[i];
}
if ( UNBIASED ) RemoveBias(Rule[NRules], InitCoeffs);
}
}
/*************************************************************************/
/* */
/* Increment counts and averages */
/* */
/*************************************************************************/
void UpdateCount(int d, CaseNo i, double *Total, double *PredErr)
/* ----------- */
{
ContValue Val, ModelVal, OldPred, NewPred;
double Wt;
DataRec C;
C = Case[i];
Val = Class(C);
Wt = CWeight(C);
Total[d] += Wt;
CPredVal[i] = ModelVal = RawLinModel(Model, C);
if ( ModelVal < Floor ) ModelVal = Floor;
else
if ( ModelVal > Ceiling ) ModelVal = Ceiling;
if ( d )
{
OldPred = PredSum(C) / PredCount(C);
NewPred = (PredSum(C) + ModelVal) / (PredCount(C) + 1);
PredErr[d] += Wt * (fabs(Val - OldPred) - fabs(Val - NewPred));
}
else
{
PredErr[0] += Wt * fabs(Val - ModelVal);
}
}
/*************************************************************************/
/* */
/* Change Fail0, Fail1, and FailMany. */
/* */
/* If Bestd has not been set, initialise the lists; otherwise */
/* record the changes for deleting condition Bestd and reduce */
/* LocalNFail for cases that do not satisfy condition Bestd */
/* */
/*************************************************************************/
void ProcessLists()
/* ------------ */
{
CaseNo i, iNext, *Prev;
int d;
if ( ! Bestd )
{
/* Initialise the lists */
Fail0 = Fail1 = FailMany = -1;
ForEach(d, 0, NCond)
{
Total[d] = PredErr[d] = 0;
}
ForEach(i, 0, MaxCase)
{
if ( ! LocalNFail[i] )
{
UpdateCount(0, i, Total, PredErr);
AddToList(&Fail0, i);
}
else
if ( LocalNFail[i] == 1 )
{
d = SingleFail(i);
UpdateCount(d, i, Total, PredErr);
AddToList(&Fail1, i);
}
else
{
AddToList(&FailMany, i);
}
}
}
else
{
/* Promote cases from Fail1 to Fail0 */
Prev = &Fail1;
for ( i = Fail1 ; i >= 0 ; )
{
iNext = Succ[i];
if ( CondFailedBy[Bestd][i] )
{
LocalNFail[i] = 0;
UpdateCount(0, i, Total, PredErr);
DeleteFromList(Prev, i);
AddToList(&Fail0, i);
}
else
{
Prev = &Succ[i];
}
i = iNext;
}
/* Check cases in FailMany */
Prev = &FailMany;
for ( i = FailMany ; i >= 0 ; )
{
iNext = Succ[i];
if ( CondFailedBy[Bestd][i] && --LocalNFail[i] == 1 )
{
d = SingleFail(i);
UpdateCount(d, i, Total, PredErr);
DeleteFromList(Prev, i);
AddToList(&Fail1, i);
}
else
{
Prev = &Succ[i];
}
i = iNext;
}
}
}
/*************************************************************************/
/* */
/* Add case to list whose first case is *List */
/* */
/*************************************************************************/
void AddToList(CaseNo *List, CaseNo N)
/* --------- */
{
Succ[N] = *List;
*List = N;
}
/*************************************************************************/
/* */
/* Delete case from list where previous case is *Before */
/* */
/*************************************************************************/
void DeleteFromList(CaseNo *Before, CaseNo N)
/* -------------- */
{
*Before = Succ[N];
}
/*************************************************************************/
/* */
/* Find single condition failed by a case */
/* */
/*************************************************************************/
int SingleFail(CaseNo i)
/* ---------- */
{
int d;
ForEach(d, 1, NCond)
{
if ( ! Deleted[d] && CondFailedBy[d][i] ) return d;
}
return 0;
}
/*************************************************************************/
/* */
/* Make sure rule is approximately unbiased */
/* */
/*************************************************************************/
void RemoveBias(CRule R, int Coeffs)
/* ---------- */
{
double Wt, TotErr=0, TotWt=0, TotAbsErr=-1, Bias=0, LastBias, New;
CaseNo i;
/* Computer initial bias and total weight */
for ( i = Fail0 ; i >= 0 ; )
{
Wt = CWeight(Case[i]);
TotWt += Wt;
New = ( CPredVal[i] < R->LoLim ? R->LoLim :
CPredVal[i] > R->HiLim ? R->HiLim : CPredVal[i] );
TotErr += Wt * (New - Class(Case[i]));
i = Succ[i];
}
Bias = TotErr / TotWt;
/* Iteratively remove bias. This cannot be done in a single step
because range truncation can affect the mean residual */
while ( fabs((LastBias = Bias)) >= 0.5 * AttUnit[0] )
{
R->Rhs[0] -= Bias;
TotErr = TotAbsErr = 0;
for ( i = Fail0 ; i >= 0 ; )
{
Wt = CWeight(Case[i]);
/* Remove previous bias from raw value */
CPredVal[i] -= Bias;
/* Compute new range-limited value */
New = ( CPredVal[i] < R->LoLim ? R->LoLim :
CPredVal[i] > R->HiLim ? R->HiLim : CPredVal[i] );
TotErr += Wt * (New - Class(Case[i]));
TotAbsErr += Wt * fabs(New - Class(Case[i]));
i = Succ[i];
}
Bias = TotErr / TotWt;
Verbosity(1, printf("bias %.5f (%.2f / %.2f)\n", Bias, TotErr, TotWt))
if ( fabs(Bias) >= fabs(LastBias) )
{
break;
}
}
/* If any changes, reset the intercept precision and estimated error */
if ( TotAbsErr >= 0 )
{
R->Rhs[0] = rint(R->Rhs[0] / AttUnit[0]) * AttUnit[0];
R->EstErr = EstimateErr(TotAbsErr / TotWt, R->Cover, Coeffs);
}
}
/*************************************************************************/
/* */
/* Sort rules by mean value */
/* */
/*************************************************************************/
void OrderRules(void)
/* ---------- */
{
RuleNo r, nr, Next;
CRule Hold;
ForEach(r, 1, NRules)
{
Next = r;
ForEach(nr, r+1, NRules)
{
if ( Rule[nr]->Mean < Rule[Next]->Mean )
{
Next = nr;
}
}
Rule[Next]->RNo = r;
if ( Next != r )
{
Hold = Rule[r];
Rule[r] = Rule[Next];
Rule[Next] = Hold;
}
}
}
/*************************************************************************/
/* */
/* Free all data allocated for forming rules */
/* */
/*************************************************************************/
void FreeFormRuleData(void)
/* ---------------- */
{
if ( ! CondFailedBy ) return;
FreeVector((void **) CondFailedBy, 0, MaxDepth+1); CondFailedBy = Nil;
Free(Stack); Stack = Nil;
Free(Deleted); Deleted = Nil;
Free(Total); Total = Nil;
Free(PredErr); PredErr = Nil;
Free(NFail); NFail = Nil;
Free(LocalNFail); LocalNFail = Nil;
Free(Succ); Succ = Nil;
Free(CPredVal); CPredVal = Nil;
}
formtree.c 0000664 0000764 0000764 00000034324 11501534415 012764 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Central tree-forming algorithm */
/* ------------------------------ */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
/*************************************************************************/
/* */
/* Allocate space for tree tables */
/* */
/*************************************************************************/
void InitialiseEnvData(void)
/* ----------------- */
{
DiscrValue v;
Attribute Att;
SRec = Alloc(MaxCase+1, SortRec);
/* Set up environment */
GEnv.LocalModel = AllocZero(MaxAtt+1, double);
GEnv.ValFreq = Alloc(MaxDiscrVal+1, double);
GEnv.ValSum = Alloc(MaxDiscrVal+1, double);
GEnv.ValSumSq = Alloc(MaxDiscrVal+1, double);
GEnv.Gain = AllocZero(MaxAtt+1, float);
GEnv.Bar = AllocZero(MaxAtt+1, ContValue);
/* Data for subsets */
GEnv.Left = Alloc(MaxDiscrVal+1, Boolean);
GEnv.Subset = AllocZero(MaxAtt+1, Set *);
ForEach(Att, 1, MaxAtt)
{
if ( Discrete(Att) )
{
GEnv.Subset[Att] = AllocZero(4, Set);
ForEach(v, 1, 3)
{
GEnv.Subset[Att][v] =
Alloc((MaxAttVal[Att]>>3)+1, unsigned char);
}
}
}
/* Stuff for forming linear models by matrix inversion */
GEnv.xTx = Alloc(MaxAtt+1, double *);
GEnv.A = Alloc(MaxAtt+1, double *);
ForEach(Att, 0, MaxAtt)
{
GEnv.xTx[Att] = Alloc(MaxAtt+1, double);
GEnv.A[Att] = Alloc(MaxAtt+1, double);
}
GEnv.xTy = Alloc(MaxAtt+1, double);
GEnv.B = Alloc(MaxAtt+1, double);
GEnv.Resid = Alloc(MaxCase+1, double);
GEnv.PResid = Alloc(MaxCase+1, double);
GEnv.Mean = Alloc(MaxAtt+1, double);
GEnv.Var = Alloc(MaxAtt+1, double);
GEnv.AvDev = Alloc(MaxAtt+1, double);
GEnv.ZeroCoeff = AllocZero(MaxAtt+1, Boolean);
GEnv.BestModel = Alloc(MaxAtt+1, double);
GEnv.SaveZero = Alloc(MaxAtt+1, Boolean);
GEnv.Filtered = Alloc(MaxCase+1, DataRec);
GEnv.DoNotUse = Alloc(MaxAtt+1, Boolean);
GEnv.ModelAtt = Alloc(MaxAtt+1, Attribute);
}
void FreeEnvData(void)
/* ------------ */
{
Attribute Att;
if ( ! GEnv.LocalModel ) return;
FreeUnlessNil(GEnv.LocalModel);
FreeUnlessNil(GEnv.ValFreq);
FreeUnlessNil(GEnv.ValSum);
FreeUnlessNil(GEnv.ValSumSq);
FreeUnlessNil(GEnv.Gain);
FreeUnlessNil(GEnv.Bar);
FreeUnlessNil(GEnv.Left);
ForEach(Att, 1, MaxAtt)
{
if ( Discrete(Att) )
{
FreeVector((void **) GEnv.Subset[Att], 1, 3);
}
}
FreeUnlessNil(GEnv.Subset);
FreeVector((void **) GEnv.xTx, 0, MaxAtt);
FreeVector((void **) GEnv.A, 0, MaxAtt);
FreeUnlessNil(GEnv.xTy);
FreeUnlessNil(GEnv.B);
FreeUnlessNil(GEnv.Resid);
FreeUnlessNil(GEnv.PResid);
FreeUnlessNil(GEnv.Mean);
FreeUnlessNil(GEnv.Var);
FreeUnlessNil(GEnv.AvDev);
FreeUnlessNil(GEnv.ZeroCoeff);
FreeUnlessNil(GEnv.BestModel);
FreeUnlessNil(GEnv.SaveZero);
FreeUnlessNil(GEnv.Filtered);
FreeUnlessNil(GEnv.DoNotUse);
FreeUnlessNil(GEnv.ModelAtt);
FreeUnlessNil(SRec); SRec = Nil;
}
/*************************************************************************/
/* */
/* Set global preferences for continuous attributes and find */
/* statistics for global target values. */
/* */
/* Attribute preference is determined by counting the number of */
/* times the sign of the deviations of the attribute and the */
/* target value agree. Very low or very high values are better. */
/* */
/*************************************************************************/
void FindGlobalProperties(void)
/* -------------------- */
{
int *N, *Same;
CaseNo i;
Attribute Att;
Boolean AboveAv;
ContValue Val;
double Sum=0, SumSq=0, Wt, MinCoeff;
Verbosity(1, fprintf(Of, "\nPreference values:\n"))
N = AllocZero(MaxAtt+1, int);
Same = AllocZero(MaxAtt+1, int);
ForEach(i, 0, MaxCase)
{
Wt = CWeight(Case[i]);
Sum += Wt * (Val = Class(Case[i]));
SumSq += Wt * Val * Val;
AboveAv = ( Val > AttMean[ClassAtt] );
ForEach(Att, 1, MaxAtt)
{
/* Exclude everything except continuous predictor attributes
and also discard non-applicable or unknown values */
if ( ! Skip(Att) && ! Discrete(Att) && Att != ClassAtt &&
! NotApplic(Case[i], Att) &&
(Val = CVal(Case[i], Att)) != AttMean[Att] )
{
N[Att]++;
Same[Att] += ( ( Val > AttMean[Att] ) == AboveAv );
}
}
}
GlobalMean = Sum / (MaxCase+1);
GlobalSD = SD(MaxCase+1, Sum, SumSq);
GlobalErr = AverageDev(GlobalMean, 0, MaxCase);
AttUnit[0] = 1.0 / AttPrec[ClassAtt];
ForEach(Att, 1, MaxAtt)
{
if ( Skip(Att) || Discrete(Att) || Att == ClassAtt || ! N[Att] )
{
AttPref[Att] = 0;
}
else
{
AttPref[Att] = (2 * Max(Same[Att], N[Att] - Same[Att]) - N[Att]) /
(double) (MaxCase+1);
Verbosity(1, fprintf(Of, "%6.3f %s\n", AttPref[Att], AttName[Att]))
/* Set coefficient unit for attribute so that
AttUnit[Att] * AttSD[Att] <= UNITFRAC * GlobalSD */
AttUnit[Att] = 1;
MinCoeff = MinAttCoeff(Att);
if ( MinCoeff > 0 )
{
for ( ; AttUnit[Att] > MinCoeff ; AttUnit[Att] /= 10 )
;
}
}
}
free(N);
free(Same);
}
/*************************************************************************/
/* */
/* Build a regression tree for the cases Fp through Lp */
/* */
/*************************************************************************/
void FormTree(CaseNo Fp, CaseNo Lp, int Level, Tree *Result, Tree Parent)
/* -------- */
{
CaseNo i;
ContValue BestVal, Val, MaxResid;
double Cases=0, Wt, RawSum=0, Sum=0, SumSq=0;
Attribute Att, BestAtt;
Tree Node, P;
Boolean Root;
Verbosity(1, fprintf(Of, "\n<%d> %d cases\n", Level, Lp - Fp + 1))
/* Find local model and residuals */
if ( ! (Root = (Lp - Fp >= MaxCase)) )
{
ForEach(Att, 1, MaxAtt)
{
GEnv.DoNotUse[Att] = true;
}
for ( P = Parent ; P ; P = Parent(P) )
{
GEnv.DoNotUse[P->Tested] = false;
}
AddDefAtts();
Regress(Fp, Lp, GEnv.LocalModel);
FindModelAtts(GEnv.LocalModel);
}
else
{
ForEach(Att, 0, MaxAtt)
{
GEnv.LocalModel[Att] = 0;
}
GEnv.NModelAtt = 0;
}
MaxResid = 0;
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
Val = Resid(Case[i]) =
( Root ? Class(Case[i]) :
Class(Case[i]) - LinModel(GEnv.LocalModel, Case[i]) );
Cases += Wt;
RawSum += Wt * Class(Case[i]);
Sum += Wt * Val;
SumSq += Wt * Val * Val;
if ( (Val = fabs(Val)) > MaxResid ) MaxResid = Val;
}
/* Establish root of this subtree */
*Result = Node = Leaf(Lp-Fp+1, RawSum / Cases, SD(Cases, Sum, SumSq));
if ( Root )
{
Node->Model[0] = Node->Mean;
}
else
{
memcpy(Node->Model, GEnv.LocalModel, (MaxAtt+1) * sizeof(double));
}
/* If the maximum residual is smaller than the desired model
precision, there is no need to go further */
if ( MaxResid < 0.5 * AttUnit[0] )
{
Verbosity(1, fprintf(Of, "negligible residuals\n"))
Progress(Cases);
return;
}
/* Find the attribute with maximum gain */
BestVal = -Epsilon;
BestAtt = None;
ForEach(Att, 1, MaxAtt)
{
GEnv.Gain[Att] = None;
if ( Skip(Att) || Att == ClassAtt ) continue;
if ( Discrete(Att) )
{
if ( Root || MaxAttVal[Att] > 3 || GEnv.DoNotUse[Att] )
{
EvalDiscreteAtt(Node, Att, Fp, Lp);
}
}
else
{
EvalContinuousAtt(Node, Att, Fp, Lp);
}
if ( (Val = GEnv.Gain[Att]) > -Epsilon )
{
if ( Val > BestVal ||
Val > 0.999 * BestVal && AttPref[Att] > AttPref[BestAtt] )
{
BestAtt = Att;
BestVal = Val;
}
}
}
/* Decide whether to branch or not */
if ( BestAtt == None )
{
Verbosity(1, fprintf(Of, "no sensible splits\n"))
Progress(Cases);
}
else
{
Verbosity(1,
fprintf(Of, "best attribute %s", AttName[BestAtt]);
if ( Continuous(BestAtt) )
{
fprintf(Of, " cut %.3f", GEnv.Bar[BestAtt]);
}
fprintf(Of, " gain %.3f\n", GEnv.Gain[BestAtt]);)
/* Build a node of the selected test */
if ( Discrete(BestAtt) )
{
DiscreteTest(Node, BestAtt, GEnv.Subset[BestAtt]);
}
else
{
ContinTest(Node, BestAtt, GEnv.Bar[BestAtt]);
}
/* Record parent of this node */
Parent(Node) = Parent;
/* Carry out the recursive divide-and-conquer */
Divide(Node, Fp, Lp, Level);
}
}
void AddModels(CaseNo Fp, CaseNo Lp, Tree T, Tree Parent)
/* --------- */
{
Attribute Att;
CaseNo Bp, Ep;
DiscrValue v;
Progress(1.0);
if ( ! T->Cases ) return;
if ( T->NodeType )
{
Bp = Fp;
ForEach(v, 1, T->Forks)
{
if ( T->Branch[v]->Cases )
{
Ep = Bp + T->Branch[v]->Cases - 1;
AddModels(Bp, Ep, T->Branch[v], T);
Bp = Ep+1;
}
}
}
/* Find a new model for this node */
ForEach(Att, 1, MaxAtt)
{
GEnv.DoNotUse[Att] = true;
}
while ( Parent )
{
if ( Continuous((Att = Parent->Tested)) )
{
GEnv.DoNotUse[Att] = false;
}
Parent = Parent(Parent);
}
AddSplitAtts(T);
AddDefAtts();
Regress(Fp, Lp, T->Model);
}
/*************************************************************************/
/* */
/* Form the subtrees for the given node */
/* */
/*************************************************************************/
void Divide(Tree Node, CaseNo Fp, CaseNo Lp, int Level)
/* ------ */
{
CaseNo Ep;
DiscrValue v;
#ifdef VerbOpt
CaseNo XEp[4]; /* Queue tasks in non-SMP order */
XEp[0] = Fp-1;
#endif
assert(Node->Forks < 4);
/* Recursive divide and conquer */
ForEach(v, 1, Node->Forks)
{
Ep = Group(v, Fp, Lp, Node);
#ifdef VerbOpt
XEp[v] = Ep;
#endif
if ( Fp <= Ep )
{
FormTree(Fp, Ep, Level+1, &Node->Branch[v], Node);
Fp = Ep + 1;
}
else
{
Node->Branch[v] = Leaf(0, Node->Mean, Node->SD);
}
}
}
/*************************************************************************/
/* */
/* Group together the items corresponding to branch V of a test */
/* and return the index of the last such */
/* */
/* Note: if V equals zero, group the unknown values */
/* */
/*************************************************************************/
CaseNo Group(DiscrValue V, CaseNo Fp, CaseNo Lp, Tree TestNode)
/* ----- */
{
CaseNo i;
Attribute Att;
ContValue Thresh;
Set SS;
DataRec xab;
Att = TestNode->Tested;
/* Group items on the value of attribute Att, and depending
on the type of branch */
switch ( TestNode->NodeType )
{
case BrDiscr:
ForEach(i, Fp, Lp)
{
if ( DVal(Case[i], Att) == V )
{
Swap(Fp, i);
Fp++;
}
}
break;
case BrThresh:
Thresh = TestNode->Cut;
ForEach(i, Fp, Lp)
{
if ( V == 1 ? NotApplic(Case[i], Att) :
(CVal(Case[i], Att) <= Thresh) == (V == 2) )
{
Swap(Fp, i);
Fp++;
}
}
break;
case BrSubset:
SS = TestNode->Subset[V];
ForEach(i, Fp, Lp)
{
if ( In(DVal(Case[i], Att), SS) )
{
Swap(Fp, i);
Fp++;
}
}
break;
}
return Fp - 1;
}
/*************************************************************************/
/* */
/* Add the continuous attributes that are used in a branch of */
/* the tree to those that can be used in a model */
/* */
/*************************************************************************/
void AddSplitAtts(Tree T)
/* ------------ */
{
DiscrValue v;
if ( T->NodeType )
{
if ( Continuous(T->Tested) )
{
GEnv.DoNotUse[T->Tested] = false;
}
ForEach(v, 1, T->Forks)
{
AddSplitAtts(T->Branch[v]);
}
}
}
/*************************************************************************/
/* */
/* If attribute A can be used in a model, then so can any */
/* implicitly-defined attribute that uses A */
/* */
/*************************************************************************/
void AddDefAtts(void)
/* ---------- */
{
Attribute Att;
Definition D;
int e;
ForEach(Att, 1, MaxAtt)
{
if ( GEnv.DoNotUse[Att] && (D = AttDef[Att]) )
{
for ( e = 0 ; ; e++ )
{
if ( DefOp(D[e]) == OP_ATT &&
! GEnv.DoNotUse[(long) DefSVal(D[e])] )
{
Verbosity(2,
fprintf(Of, "adding %s from %s\n",
AttName[Att], AttName[(int) DefSVal(D[e])]))
GEnv.DoNotUse[Att] = false;
break;
}
else
if ( DefOp(D[e]) == OP_END )
{
break;
}
}
}
}
}
/*************************************************************************/
/* */
/* Find attributes used in a model (speeds up LinModel) */
/* */
/*************************************************************************/
void FindModelAtts(double *Model)
/* ------------- */
{
Attribute Att;
GEnv.NModelAtt = 0;
ForEach(Att, 1, MaxAtt)
{
if ( Model[Att] )
{
GEnv.ModelAtt[++GEnv.NModelAtt] = Att;
}
}
}
getdata.c 0000664 0000764 0000764 00000036430 11501534415 012552 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Get case descriptions from data file */
/* -------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define Inc 2048
/* Alternative random number generator */
#define AltRandom drand48()
double drand48(void);
Boolean SuppressErrorMessages=false;
#define XError(a,b,c) if (! SuppressErrorMessages) Error(a,b,c)
CaseNo SampleFrom; /* file count for sampling */
/*************************************************************************/
/* */
/* Read raw cases from file with given extension. */
/* */
/* On completion, cases are stored in array Case in the form */
/* of vectors of attribute values, and MaxCase is set to the */
/* number of data cases. */
/* */
/*************************************************************************/
void GetData(FILE *Df, Boolean Train, Boolean AllowUnknownTarget)
/* ------- */
{
CaseNo CaseSpace, i;
DataRec DVec;
Boolean AnyUnknown=false, FirstIgnore=true, *AttMsg;
Attribute Att;
ContValue Val, Range;
char CVS[20];
CaseCount *Freq; /* discrete value frequencies */
CaseNo Count, WantTrain, LeftTrain, WantTest, LeftTest;
Boolean SelectTrain;
double Sum, SumSq;
DiscrValue v;
LineNo = 0;
SuppressErrorMessages = SAMPLE && ! Train;
if ( Train || ! Case )
{
MaxCase = MaxLabel = CaseSpace = 0;
Case = Alloc(1, DataRec); /* for error reporting */
}
else
{
CaseSpace = MaxCase + 1;
MaxCase++;
}
if ( SAMPLE )
{
if ( Train )
{
SampleFrom = CountData(Df);
ResetKR(KRInit); /* initialise KRandom() */
}
else
{
ResetKR(KRInit); /* restore KRandom() */
}
WantTrain = SampleFrom * SAMPLE + 0.5;
LeftTrain = SampleFrom;
WantTest = ( SAMPLE < 0.5 ? WantTrain : SampleFrom - WantTrain );
LeftTest = SampleFrom - WantTrain;
}
while ( (DVec = GetDataRec(Df, Train)) )
{
/* Check whether to include if we are sampling */
if ( SAMPLE )
{
SelectTrain = KRandom() < WantTrain / (float) LeftTrain--;
/* Include if
* Select and this is the training set
* ! Select and this is the test set and sub-select
NB: Must use different random number generator for
sub-selection since cannot disturb random number sequence */
if ( SelectTrain )
{
WantTrain--;
}
if ( SelectTrain != Train ||
( ! Train && AltRandom >= WantTest / (float) LeftTest-- ) )
{
FreeLastCase(DVec);
continue;
}
if ( ! Train )
{
WantTest--;
}
}
/* Make sure there is room for another item */
if ( MaxCase >= CaseSpace )
{
CaseSpace += Inc;
Realloc(Case, CaseSpace+1, DataRec);
}
/* Ignore cases with N/A target value (!) but possibly allow
cases with unknown target value */
if ( ! NotApplic(DVec, ClassAtt) &&
( AllowUnknownTarget || CVal(DVec, ClassAtt) != UNKNOWN ) )
{
Case[MaxCase] = DVec;
MaxCase++;
}
else
{
if ( FirstIgnore && Of )
{
fprintf(Of, ( AllowUnknownTarget ? T_IgnoreNATarget :
T_IgnoreBadTarget ));
FirstIgnore = false;
}
FreeLastCase(DVec);
}
}
fclose(Df);
MaxCase--;
if ( Of && MaxCase < 0 )
{
fprintf(Of, T_NoCases);
Goodbye(1);
}
/* Replace unknown values by means or modals */
if ( AttMean == Nil )
{
/* Find means, limits, modal values and precision */
AttMean = AllocZero(MaxAtt+1, ContValue);
AttSD = AllocZero(MaxAtt+1, ContValue);
AttMaxVal = AllocZero(MaxAtt+1, ContValue);
AttMinVal = AllocZero(MaxAtt+1, ContValue);
AttPref = AllocZero(MaxAtt+1, ContValue);
AttUnit = AllocZero(MaxAtt+1, double);
AttPrec = AllocZero(MaxAtt+1, int);
Modal = Alloc(MaxAtt+1, DiscrValue);
Freq = Alloc(MaxDiscrVal+1, CaseCount);
ForEach(Att, 1, MaxAtt)
{
if ( Exclude(Att) ) continue;
if ( Continuous(Att) )
{
Sum = SumSq = Count = 0;
ForEach(i, 0, MaxCase)
{
Val = CVal(Case[i], Att);
if ( Val != UNKNOWN && DVal(Case[i], Att) != NA )
{
Sum += Val;
SumSq += Val * Val;
Count++;
}
}
if ( Count > 1 )
{
AttPrec[Att] = 10 * FracBase(Att);
AttMean[Att] = floor((Sum / Count) * AttPrec[Att] + 0.1) /
AttPrec[Att];
AttSD[Att] = sqrt((SumSq - Sum * Sum / Count) / (Count-1))
+ 1E-6;
}
else
{
AttPrec[Att] = 0;
AttMean[Att] = 0;
AttSD[Att] = 1;
SpecialStatus[Att] |= SKIP;
}
FindLimits(Att, &AttMinVal[Att], &AttMaxVal[Att]);
}
else
{
ForEach(v, 0, MaxAttVal[Att])
{
Freq[v] = 0;
}
ForEach(i, 0, MaxCase)
{
if ( (v = DVal(Case[i], Att)) ) Freq[v]++;
}
Modal[Att] = 0; /* cannot be N/A */
ForEach(v, 2, MaxAttVal[Att])
{
if ( Freq[v] > Freq[Modal[Att]] ) Modal[Att] = v;
}
if ( ! Modal[Att] )
{
SpecialStatus[Att] |= SKIP;
}
}
}
Free(Freq); Freq = Nil;
Precision = log(AttPrec[ClassAtt] / 10.0) / log(10.0) + 0.5;
}
/* Now, the actual replacement */
AttMsg = AllocZero(MaxAtt+1, Boolean);
ForEach(i, 0, MaxCase)
{
AnyUnknown |= ReplaceUnknowns(Case[i], AttMsg);
}
if ( Of && Train && AnyUnknown )
{
fprintf(Of, T_ReplaceUnknowns);
ForEach(Att, 1, MaxAtt)
{
if ( AttMsg[Att] )
{
if ( Skip(Att) )
{
fprintf(Of, "\t`%s' -- " T_NoAppVals "\n", AttName[Att]);
continue;
}
fprintf(Of, "\t`%s' " T_By " ", AttName[Att]);
if ( Discrete(Att) )
{
fprintf(Of, "`%s'\n", AttValName[Att][Modal[Att]]);
}
else
{
CValToStr(AttMean[Att], Att, CVS);
fprintf(Of, "%s\n", CVS);
}
}
}
}
if ( Train )
{
/* Find global range and set lowest and highest predicted value */
Range = AttMaxVal[ClassAtt] - AttMinVal[ClassAtt];
Ceiling = AttMaxVal[ClassAtt] + EXTRAP * Range;
if ( AttMaxVal[ClassAtt] <= 0 && Ceiling > 0 ) Ceiling = 0;
Floor = AttMinVal[ClassAtt] - EXTRAP * Range;
if ( AttMinVal[ClassAtt] >= 0 && Floor < 0 ) Floor = 0;
}
Free(AttMsg); AttMsg = Nil;
}
/*************************************************************************/
/* */
/* Replace any unknown values in a case. If AttMsg is not Nil, */
/* note the attributes involved. */
/* */
/* Note: From release 1.13, N/A values are not replaced */
/* */
/*************************************************************************/
Boolean ReplaceUnknowns(DataRec Case, Boolean *AttMsg)
/* --------------- */
{
Attribute Att;
Boolean Replaced=false;
ForEach(Att, 1, MaxAtt)
{
if ( Skip(Att) || Att == ClassAtt ) continue;
if ( Discrete(Att) && ! DVal(Case, Att) )
{
DVal(Case, Att) = Modal[Att];
if ( AttMsg ) AttMsg[Att] = Replaced = true;
}
else
if ( Continuous(Att) && CVal(Case, Att) == UNKNOWN )
{
CVal(Case, Att) = AttMean[Att];
if ( AttMsg ) AttMsg[Att] = Replaced = true;
}
}
Class(Case) = CVal(Case, ClassAtt);
return Replaced;
}
/*************************************************************************/
/* */
/* Read a raw case from file Df. */
/* */
/* For each attribute, read the attribute value from the file. */
/* If it is a discrete valued attribute, find the associated no. */
/* of this attribute value (if the value is unknown this is 0). */
/* */
/* Returns the DataRec of the case (i.e. the array of attribute */
/* values). */
/* */
/*************************************************************************/
DataRec GetDataRec(FILE *Df, Boolean Train)
/* ---------- */
{
Attribute Att;
char Name[1000], *EndVal;
int Dv, Chars;
ContValue Cv;
DataRec DVec;
Boolean FirstValue=true;
if ( ReadName(Df, Name, 1000, '\00') )
{
Case[MaxCase] = DVec = NewCase();
ForEach(Att, 1, MaxAtt)
{
if ( AttDef[Att] )
{
DVec[Att] = EvaluateDef(AttDef[Att], DVec);
if ( Continuous(Att) )
{
CheckValue(DVec, Att);
}
continue;
}
/* Get the attribute value if don't already have it */
if ( ! FirstValue && ! ReadName(Df, Name, 1000, '\00') )
{
XError(EOFINATT, AttName[Att], "");
FreeLastCase(DVec);
return Nil;
}
FirstValue = false;
if ( Exclude(Att) )
{
if ( Att == LabelAtt )
{
/* Record the value as a string */
SVal(DVec,Att) = StoreIVal(Name);
}
}
else
if ( ! strcmp(Name, "?") )
{
/* Unknown value */
if ( Continuous(Att) )
{
CVal(DVec, Att) = UNKNOWN;
}
else
{
DVal(DVec, Att) = 0;
}
}
else
if ( ! strcmp(Name, "N/A") )
{
/* Non-applicable value */
DVal(DVec, Att) = NA;
}
else
if ( Discrete(Att) )
{
Dv = Which(Name, AttValName[Att], 1, MaxAttVal[Att]);
if ( ! Dv )
{
if ( StatBit(Att, DISCRETE) )
{
if ( Train || XVAL )
{
/* Add value to list */
if ( MaxAttVal[Att] >= (long) AttValName[Att][0] )
{
XError(TOOMANYVALS, AttName[Att],
(char *) AttValName[Att][0] - 1);
Dv = MaxAttVal[Att];
}
else
{
Dv = ++MaxAttVal[Att];
AttValName[Att][Dv] = strdup(Name);
AttValName[Att][Dv+1] = ""; /* no free */
}
if ( Dv > MaxDiscrVal )
{
MaxDiscrVal = Dv;
}
}
else
{
/* Set value to "" */
Dv = MaxAttVal[Att] + 1;
}
}
else
{
XError(BADATTVAL, AttName[Att], Name);
}
}
DVal(DVec, Att) = Dv;
}
else
{
/* Continuous value */
if ( TStampVal(Att) )
{
Cv = TStampToMins(Name);
if ( Cv >= 1E9 ) /* long time in future */
{
XError(BADTSTMP, AttName[Att], Name);
Cv = UNKNOWN;
}
}
else
if ( DateVal(Att) )
{
Cv = DateToDay(Name);
if ( Cv < 1 )
{
XError(BADDATE, AttName[Att], Name);
Cv = UNKNOWN;
}
}
else
if ( TimeVal(Att) )
{
Cv = TimeToSecs(Name);
if ( Cv < 0 )
{
XError(BADTIME, AttName[Att], Name);
Cv = UNKNOWN;
}
}
else
{
Cv = strtod(Name, &EndVal);
if ( EndVal == Name || *EndVal != '\0' )
{
XError(BADATTVAL, AttName[Att], Name);
Cv = UNKNOWN;
}
}
CVal(DVec, Att) = Cv;
CheckValue(DVec, Att);
}
}
Class(DVec) = CVal(DVec, ClassAtt);
if ( LabelAtt &&
(Chars = strlen(IgnoredVals + SVal(DVec, LabelAtt))) > MaxLabel )
{
MaxLabel = Chars;
}
return DVec;
}
else
{
return Nil;
}
}
/*************************************************************************/
/* */
/* Count cases in data file */
/* */
/*************************************************************************/
CaseNo CountData(FILE *Df)
/* --------- */
{
char Last=',';
int Count=0, Next;
while ( true )
{
if ( (Next = getc(Df)) == EOF )
{
if ( Last != ',' ) Count++;
rewind(Df);
return Count;
}
if ( Next == '|' )
{
while ( (Next = getc(Df)) != '\n' )
;
}
if ( Next == '\n' )
{
if ( Last != ',' ) Count++;
Last = ',';
}
else
if ( Next == '\\' )
{
/* Skip escaped character */
getc(Df);
}
else
if ( Next != '\t' && Next != ' ' )
{
Last = Next;
}
}
}
/*************************************************************************/
/* */
/* Store a label or ignored value in IValStore */
/* */
/*************************************************************************/
int StoreIVal(String S)
/* --------- */
{
int StartIx, Length;
if ( (Length=strlen(S) + 1) + IValsOffset > IValsSize )
{
if ( IgnoredVals )
{
Realloc(IgnoredVals, IValsSize += 32768, char);
}
else
{
IValsSize = 32768;
IValsOffset = 0;
IgnoredVals = Alloc(IValsSize, char);
}
}
StartIx = IValsOffset;
strcpy(IgnoredVals + StartIx, S);
IValsOffset += Length;
return StartIx;
}
/*************************************************************************/
/* */
/* Free case description space */
/* */
/*************************************************************************/
void FreeData(DataRec *Case)
/* -------- */
{
/* Do not free entire data block when freeing instances.
Otherwise cannot free excluded attributes in real data */
if ( Case != Instance )
{
FreeCases();
}
FreeUnlessNil(IgnoredVals); IgnoredVals = Nil;
IValsSize = 0;
Free(Case);
}
/*************************************************************************/
/* */
/* Check for bad continuous value */
/* */
/*************************************************************************/
void CheckValue(DataRec DVec, Attribute Att)
/* ---------- */
{
ContValue Cv;
Cv = CVal(DVec, Att);
if ( ! finite(Cv) )
{
Error(BADNUMBER, AttName[Att], "");
CVal(DVec, Att) = UNKNOWN;
DVal(DVec, Att) = 0;
}
}
/*************************************************************************/
/* */
/* Find maximum and minimum values for a continuous attribute */
/* */
/*************************************************************************/
void FindLimits(Attribute Att, ContValue *Min, ContValue *Max)
/* ---------- */
{
ContValue Val;
CaseNo i;
*Max = -1E38;
*Min = 1E38;
ForEach(i, 0, MaxCase)
{
if ( (Val = CVal(Case[i], Att) ) == UNKNOWN ) continue;
if ( Val > *Max ) *Max = Val;
if ( Val < *Min ) *Min = Val;
}
}
getnames.c 0000664 0000764 0000764 00000032731 11501534415 012744 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Get names of attributes and attribute values */
/* -------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#include
#include
#define MAXLINEBUFFER 10000
int Delimiter;
char LineBuffer[MAXLINEBUFFER], *LBp=LineBuffer;
/*************************************************************************/
/* */
/* Read a name from file f into string s, setting Delimiter. */
/* */
/* - Embedded periods are permitted, but periods followed by space */
/* characters act as delimiters. */
/* - Embedded spaces are permitted, but multiple spaces are */
/* replaced by a single space. */
/* - Any character can be escaped by '\'. */
/* - The remainder of a line following '|' is ignored. */
/* */
/* Colons are sometimes delimiters depending on ColonOpt */
/* */
/*************************************************************************/
Boolean ReadName(FILE *f, String s, int n, char ColonOpt)
/* -------- */
{
register char *Sp=s;
register int c;
char Msg[2];
/* Skip to first non-space character */
while ( (c = InChar(f)) == '|' || Space(c) )
{
if ( c == '|' ) SkipComment;
}
/* Return false if no names to read */
if ( c == EOF )
{
Delimiter = EOF;
return false;
}
/* Read in characters up to the next delimiter */
while ( c != ColonOpt && c != ',' && c != '\n' && c != '|' && c != EOF )
{
if ( --n <= 0 )
{
if ( Of ) Error(LONGNAME, "", "");
}
if ( c == '.' )
{
if ( (c = InChar(f)) == '|' || Space(c) || c == EOF ) break;
*Sp++ = '.';
continue;
}
if ( c == '\\' )
{
c = InChar(f);
}
if ( Space(c) )
{
*Sp++ = ' ';
while ( ( c = InChar(f) ) == ' ' || c == '\t' )
;
}
else
{
*Sp++ = c;
c = InChar(f);
}
}
if ( c == '|' ) SkipComment;
Delimiter = c;
/* Special case for ':=' */
if ( Delimiter == ':' )
{
if ( *LBp == '=' )
{
Delimiter = '=';
LBp++;
}
}
/* Strip trailing spaces */
while ( Sp > s && Space(*(Sp-1)) ) Sp--;
if ( Sp == s )
{
Msg[0] = ( Space(c) ? '.' : c );
Msg[1] = '\00';
Error(MISSNAME, Fn, Msg);
}
*Sp++ = '\0';
return true;
}
/*************************************************************************/
/* */
/* Read names of attributes and legal attribute values. */
/* On completion, names are stored in: */
/* AttName - attribute names */
/* AttValName - attribute value names */
/* with: */
/* MaxAttVal - number of values for each attribute */
/* */
/* Other global variables set are: */
/* MaxAtt - maximum attribute number */
/* MaxDiscrVal - maximum discrete values for an attribute */
/* */
/*************************************************************************/
void GetNames(FILE *Nf)
/* -------- */
{
char Buffer[1000]="", *Target;
int AttCeiling=100;
Attribute Att;
ErrMsgs = AttExIn = LineNo = 0;
LBp = LineBuffer;
*LBp = 0;
/* Get name of dependent att */
ReadName(Nf, Buffer, 1000, ':');
Target = strdup(Buffer);
/* Get attribute and attribute value names from names file */
AttName = AllocZero(AttCeiling, String);
MaxAttVal = AllocZero(AttCeiling, DiscrValue);
AttValName = AllocZero(AttCeiling, String *);
SpecialStatus = AllocZero(AttCeiling, char);
AttDef = AllocZero(AttCeiling, Definition);
AttDefUses = AllocZero(AttCeiling, Attribute *);
MaxAtt = LabelAtt = CWtAtt = 0;
while ( ReadName(Nf, Buffer, 1000, ':') )
{
if ( Delimiter != ':' && Delimiter != '=' )
{
Error(BADATTNAME, Buffer, "");
}
/* Check for include/exclude instruction */
if ( ( *Buffer == 'a' || *Buffer == 'A' ) &&
! memcmp(Buffer+1, "ttributes ", 10) &&
! memcmp(Buffer+strlen(Buffer)-6, "cluded", 6) )
{
AttExIn = ( ! memcmp(Buffer+strlen(Buffer)-8, "in", 2) ? 1 : -1 );
if ( AttExIn == 1 )
{
ForEach(Att, 1, MaxAtt)
{
SpecialStatus[Att] |= SKIP;
}
}
while ( ReadName(Nf, Buffer, 1000, ':') )
{
Att = Which(Buffer, AttName, 1, MaxAtt);
if ( ! Att )
{
Error(UNKNOWNATT, Buffer, Nil);
}
else
if ( AttExIn == 1 )
{
SpecialStatus[Att] -= SKIP;
}
else
{
SpecialStatus[Att] |= SKIP;
}
}
break;
}
if ( Which(Buffer, AttName, 1, MaxAtt) > 0 )
{
Error(DUPATTNAME, Buffer, Nil);
}
if ( ++MaxAtt >= AttCeiling )
{
AttCeiling += 100;
Realloc(AttName, AttCeiling, String);
Realloc(MaxAttVal, AttCeiling, DiscrValue);
Realloc(AttValName, AttCeiling, String *);
Realloc(SpecialStatus, AttCeiling, char);
Realloc(AttDef, AttCeiling, Definition);
Realloc(AttDefUses, AttCeiling, Attribute *);
}
AttName[MaxAtt] = strdup(Buffer);
SpecialStatus[MaxAtt] = 0;
AttDef[MaxAtt] = Nil;
MaxAttVal[MaxAtt] = 0;
AttDefUses[MaxAtt] = Nil;
if ( Delimiter == '=' )
{
ImplicitAtt(Nf);
ListAttsUsed();
}
else
{
ExplicitAtt(Nf);
}
/* Check for case weight attribute, which must be type continuous */
if ( ! strcmp(AttName[MaxAtt], "case weight") )
{
CWtAtt = MaxAtt;
if ( ! Continuous(CWtAtt) )
{
Error(CWTATTERR, "", "");
}
}
}
ClassAtt = Which(Target, AttName, 1, MaxAtt);
/* Make sure not excluding class attribute */
if ( Skip(ClassAtt) ) SpecialStatus[ClassAtt] -= SKIP;
/* Class attribute must be defined and must be continuous */
if ( ClassAtt <= 0 )
{
Error(NOTARGET, Target, "");
}
else
if ( MaxAttVal[ClassAtt] > 0 ||
StatBit(ClassAtt, DISCRETE|DATEVAL|STIMEVAL|EXCLUDE) )
{
Error(BADTARGET, Target, "");
}
/* Ignore case weight attribute if it is excluded; otherwise,
it cannot be used in models */
if ( CWtAtt )
{
if ( Skip(CWtAtt) )
{
CWtAtt = 0;
}
else
{
SpecialStatus[CWtAtt] |= SKIP;
}
}
fclose(Nf);
Free(Target);
if ( ErrMsgs > 0 ) Goodbye(1);
}
/*************************************************************************/
/* */
/* Continuous or discrete attribute */
/* */
/*************************************************************************/
void ExplicitAtt(FILE *Nf)
/* ----------- */
{
char Buffer[1000]="", *p;
DiscrValue v;
int ValCeiling=100, BaseYear;
time_t clock;
/* Read attribute type or first discrete value */
if ( ! ( ReadName(Nf, Buffer, 1000, ':') ) )
{
Error(EOFINATT, AttName[MaxAtt], "");
}
MaxAttVal[MaxAtt] = 0;
if ( Delimiter != ',' )
{
/* Typed attribute */
if ( ! strcmp(Buffer, "continuous") )
{
}
else
if ( ! strcmp(Buffer, "timestamp") )
{
SpecialStatus[MaxAtt] = TSTMPVAL;
/* Set the base date if not done already */
if ( ! TSBase )
{
clock = time(0);
BaseYear = gmtime(&clock)->tm_year + 1900;
SetTSBase(BaseYear);
}
}
else
if ( ! strcmp(Buffer, "date") )
{
SpecialStatus[MaxAtt] = DATEVAL;
}
else
if ( ! strcmp(Buffer, "time") )
{
SpecialStatus[MaxAtt] = STIMEVAL;
}
else
if ( ! memcmp(Buffer, "discrete", 8) )
{
SpecialStatus[MaxAtt] = DISCRETE;
/* Read max values and reserve space */
v = atoi(&Buffer[8]);
if ( v < 2 )
{
Error(BADDISCRETE, AttName[MaxAtt], "");
}
AttValName[MaxAtt] = Alloc(v+3, String);
AttValName[MaxAtt][0] = (char *) (long) v+1;
AttValName[MaxAtt][(MaxAttVal[MaxAtt]=1)] = strdup("N/A");
}
else
if ( ! strcmp(Buffer, "ignore") )
{
SpecialStatus[MaxAtt] = EXCLUDE;
}
else
if ( ! strcmp(Buffer, "label") )
{
LabelAtt = MaxAtt;
SpecialStatus[MaxAtt] = EXCLUDE;
}
else
{
/* Cannot have only one discrete value for an attribute */
Error(SINGLEATTVAL, AttName[MaxAtt], Buffer);
}
}
else
{
/* Discrete attribute with explicit values */
AttValName[MaxAtt] = AllocZero(ValCeiling, String);
/* Add "N/A" */
AttValName[MaxAtt][(MaxAttVal[MaxAtt]=1)] = strdup("N/A");
p = Buffer;
/* Special check for ordered attribute */
if ( ! memcmp(Buffer, "[ordered]", 9) )
{
SpecialStatus[MaxAtt] = ORDERED;
for ( p = Buffer+9 ; Space(*p) ; p++ )
;
}
/* Record first real explicit value */
AttValName[MaxAtt][++MaxAttVal[MaxAtt]] = strdup(p);
/* Record remaining values */
do
{
if ( ! ( ReadName(Nf, Buffer, 1000, ':') ) )
{
Error(EOFINATT, AttName[MaxAtt], "");
}
if ( ++MaxAttVal[MaxAtt] >= ValCeiling )
{
ValCeiling += 100;
Realloc(AttValName[MaxAtt], ValCeiling, String);
}
AttValName[MaxAtt][MaxAttVal[MaxAtt]] = strdup(Buffer);
}
while ( Delimiter == ',' );
/* Cancel ordered status if <3 real values */
if ( Ordered(MaxAtt) && MaxAttVal[MaxAtt] <= 3 )
{
SpecialStatus[MaxAtt] = 0;
}
if ( MaxAttVal[MaxAtt] > MaxDiscrVal ) MaxDiscrVal = MaxAttVal[MaxAtt];
}
}
/*************************************************************************/
/* */
/* Locate value Val in List[First] to List[Last] */
/* */
/*************************************************************************/
int Which(String Val, String *List, int First, int Last)
/* ----- */
{
int n=First;
while ( n <= Last && strcmp(Val, List[n]) ) n++;
return ( n <= Last ? n : First-1 );
}
/*************************************************************************/
/* */
/* Build list of attributes used in current attribute definition */
/* AttDefUses[Att][0] = number of atts used */
/* AttDefUses[Att][1..] are the atts */
/* */
/*************************************************************************/
void ListAttsUsed(void)
/* ------------ */
{
Attribute Att;
Boolean *DefUses;
Definition D;
int e, NUsed=0;
DefUses = AllocZero(MaxAtt+1, Boolean);
D = AttDef[MaxAtt];
for ( e = 0 ; ; e++ )
{
if ( DefOp(D[e]) == OP_ATT )
{
Att = (Attribute) DefSVal(D[e]);
if ( ! DefUses[Att] )
{
DefUses[Att] = true;
NUsed++;
}
}
else
if ( DefOp(D[e]) == OP_END )
{
break;
}
}
if ( NUsed )
{
AttDefUses[MaxAtt] = Alloc(NUsed+1, Attribute);
AttDefUses[MaxAtt][0] = NUsed;
NUsed=0;
ForEach(Att, 1, MaxAtt-1)
{
if ( DefUses[Att] )
{
AttDefUses[MaxAtt][++NUsed] = Att;
}
}
}
Free(DefUses);
}
/*************************************************************************/
/* */
/* Free global names data */
/* */
/*************************************************************************/
void FreeNamesData(void)
/* ------------- */
{
Attribute a, t;
FreeVector((void **) AttName, 1, MaxAtt); AttName = Nil;
ForEach(a, 1, MaxAtt)
{
if ( a != ClassAtt && Discrete(a) )
{
FreeVector((void **) AttValName[a], 1, MaxAttVal[a]);
}
}
FreeUnlessNil(AttValName); AttValName = Nil;
/* Definitions (if any) */
if ( AttDef )
{
ForEach(a, 1, MaxAtt)
{
if ( AttDef[a] )
{
for ( t = 0 ; DefOp(AttDef[a][t]) != OP_END ; t++ )
{
if ( DefOp(AttDef[a][t]) == OP_STR )
{
Free(DefSVal(AttDef[a][t]));
}
}
Free(AttDef[a]);
Free(AttDefUses[a]);
}
}
Free(AttDef); AttDef = Nil;
Free(AttDefUses); AttDefUses = Nil;
}
FreeUnlessNil(MaxAttVal); MaxAttVal = Nil;
FreeUnlessNil(SpecialStatus); SpecialStatus = Nil;
FreeUnlessNil(AttMean); AttMean = Nil;
FreeUnlessNil(AttSD); AttSD = Nil;
FreeUnlessNil(AttMaxVal); AttMaxVal = Nil;
FreeUnlessNil(AttMinVal); AttMinVal = Nil;
FreeUnlessNil(AttPrec); AttPrec = Nil;
FreeUnlessNil(Modal); Modal = Nil;
FreeUnlessNil(AttUnit); AttUnit = Nil;
FreeUnlessNil(AttPref); AttPref = Nil;
}
/*************************************************************************/
/* */
/* Read next char keeping track of line numbers */
/* */
/*************************************************************************/
int InChar(FILE *f)
/* ------ */
{
if ( ! *LBp )
{
LBp = LineBuffer;
if ( ! fgets(LineBuffer, MAXLINEBUFFER, f) )
{
LineBuffer[0] = '\00';
return EOF;
}
LineNo++;
}
return (int) *LBp++;
}
global.c 0000664 0000764 0000764 00000013205 11501534415 012374 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* General data for Cubist */
/* ----------------------- */
/* */
/*************************************************************************/
#include "defns.i"
Attribute
ClassAtt=0, /* attribute to use as class */
LabelAtt=0, /* attribute to use as case ID */
CWtAtt=0; /* attribute to use as case weight */
char *IgnoredVals=0; /* values of labels and atts marked ignore */
int IValsSize=0, /* size of above */
IValsOffset=0; /* index of first free char */
int MaxAtt, /* max att number */
MaxDiscrVal=3, /* max discrete values for any att */
Precision, /* decimal places for target */
MaxLabel=0, /* max characters in case label */
LineNo=0, /* input line number */
ErrMsgs=0, /* errors found */
AttExIn=0, /* attribute exclusions/inclusions */
TSBase=0; /* base day for time stamps */
CaseNo MaxCase=-1; /* max data item number */
DataRec *Case; /* data items */
DataRec *SaveCase=Nil, /* original case order for better caching */
*Blocked=Nil; /* cross-validation blocks */
CaseNo SaveMaxCase; /* original number of cases */
DiscrValue
*MaxAttVal=Nil, /* number of values for each att */
*Modal=Nil; /* most frequent value for discr att */
char *SpecialStatus=Nil; /* special att treatment */
Definition
*AttDef=Nil; /* definitions of implicit atts */
Attribute
**AttDefUses=Nil; /* list of attributes used by definition */
String *AttName=Nil, /* att names */
**AttValName=Nil; /* att value names */
FILE *Of=0; /* output file */
String FileStem="undefined";
ContValue
*AttMean=Nil, /* means of att values */
*AttSD=Nil, /* std dev ditto */
*AttMaxVal=Nil, /* maximum value in training data */
*AttMinVal=Nil, /* minimum ditto */
*AttPref=Nil, /* preference value */
Ceiling, /* max allowable global prediction */
Floor, /* min allowable global prediction */
AvCWt; /* average case weight */
float ErrReduction=1; /* benefit of committee model */
double *AttUnit=Nil; /* units in which attribute reported */
int *AttPrec=Nil; /* Attribute precision */
DataRec *Instance=Nil, /* training cases */
Ref[2]; /* reference points */
CaseNo MaxInstance=-1; /* highest instance */
Index KDTree=Nil; /* index of same */
NNEnvRec
GNNEnv; /* global NN environment */
float *RSPredVal=Nil; /* tabulated RS predictions */
/*************************************************************************/
/* */
/* Global data for Cubist used for building model trees */
/* ---------------------------------------------------- */
/* */
/*************************************************************************/
EnvRec GEnv; /* global environment */
Tree TempMT=Nil; /* intermediate model tree */
SortRec *SRec=Nil; /* cache for sorting */
float GlobalMean, /* mean of entire training set */
GlobalSD, /* std dev of entire training set */
GlobalErr; /* av abs error over training set */
char Fn[512]; /* file name */
FILE *Mf=0, /* file for saving models */
*Pf=0; /* file for predicted test values */
/*************************************************************************/
/* */
/* Global data for constructing and applying rules */
/* ----------------------------------------------- */
/* */
/*************************************************************************/
CRule *Rule=Nil; /* current rules */
RuleNo NRules; /* number of rules */
int RuleSpace; /* space currently allocated for rules */
RRuleSet
*Cttee=Nil;
/*************************************************************************/
/* */
/* Global parameters for Cubist */
/* ---------------------------- */
/* */
/*************************************************************************/
int VERBOSITY=0, /* verbosity level (0 = none) */
FOLDS=10, /* cross-validation folds */
NN=0, /* nearest neighbors to use */
MEMBERS=1; /* members in committee */
float MAXD; /* max distance for close neighbors */
Boolean XVAL=0, /* true if perform crossvalidation */
CHOOSEMODE=0, /* choose whether to use instances */
USEINSTANCES=0, /* using instances */
UNBIASED=0; /* correct any rule bias */
float SAMPLE=0.0; /* sample training proportion */
int KRInit=0; /* KRandom initializer for SAMPLE */
Boolean LOCK=false; /* true if sample locked */
CaseCount
MINITEMS; /* min rule coverage */
int MAXRULES=100; /* max number of rules */
float EXTRAP=0.1; /* allowed extrapolation from models */
implicitatt.c 0000664 0000764 0000764 00000055304 11501534415 013465 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Routines to handle implicitly-defined attributes */
/* ------------------------------------------------ */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#include
char *Buff; /* buffer for input characters */
int BuffSize, BN; /* size and index of next character */
EltRec *TStack; /* expression stack model */
int TStackSize, TSN; /* size of stack and index of next entry */
int DefSize, DN; /* size of definition and next element */
Boolean PreviousError; /* to avoid parasitic errors */
AttValue _UNK, /* quasi-constant for unknown value */
_NA; /* ditto for not applicable */
#define FailSyn(Msg) {DefSyntaxError(Msg); return false;}
#define FailSem(Msg) {DefSemanticsError(Fi, Msg, OpCode); return false;}
typedef union _xstack_elt
{
DiscrValue _discr_val;
ContValue _cont_val;
String _string_val;
}
XStackElt;
#define cval _cont_val
#define sval _string_val
#define dval _discr_val
#ifdef CUBIST
#define Unknown(c,a) (! DVal(c,a))
#endif
/*************************************************************************/
/* */
/* A definition is handled in two stages: */
/* - The definition is read (up to a line ending with a period) */
/* replacing multiple whitespace characters with one space */
/* - The definition is then read (using a recursive descent */
/* parser), building up a reverse polish expression */
/* Syntax and semantics errors are flagged */
/* */
/*************************************************************************/
void ImplicitAtt(FILE *Nf)
/* ----------- */
{
#ifdef CUBIST
_UNK.cval = UNKNOWN;
#else
_UNK.dval = UNKNOWN;
#endif
_NA.dval = NA;
/* Get definition as a string in Buff */
ReadDefinition(Nf);
PreviousError = false;
BN = 0;
/* Allocate initial stack and attribute definition */
TStack = Alloc(TStackSize=50, EltRec);
TSN = 0;
AttDef[MaxAtt] = Alloc(DefSize = 100, DefElt);
DN = 0;
/* Parse Buff as an expression terminated by a period */
Expression();
if ( ! Find(".") ) DefSyntaxError("'.' ending definition");
/* Final check -- defined attribute must not be of type String */
if ( ! PreviousError )
{
if ( DN == 1 && DefOp(AttDef[MaxAtt][0]) == OP_ATT &&
strcmp(AttName[MaxAtt], "case weight") )
{
Error(SAMEATT, AttName[ (long) DefSVal(AttDef[MaxAtt][0]) ], Nil);
}
if ( TStack[0].Type == 'B' )
{
/* Defined attributes should never have a value N/A */
MaxAttVal[MaxAtt] = 3;
AttValName[MaxAtt] = AllocZero(4, String);
AttValName[MaxAtt][1] = strdup("??");
AttValName[MaxAtt][2] = strdup("t");
AttValName[MaxAtt][3] = strdup("f");
}
else
{
MaxAttVal[MaxAtt] = 0;
}
}
if ( PreviousError )
{
DN = 0;
SpecialStatus[MaxAtt] = EXCLUDE;
}
/* Write a terminating marker */
DefOp(AttDef[MaxAtt][DN]) = OP_END;
Free(Buff);
Free(TStack);
}
/*************************************************************************/
/* */
/* Read the text of a definition. Skip comments, collapse */
/* multiple whitespace characters. */
/* */
/*************************************************************************/
void ReadDefinition(FILE *f)
/* -------------- */
{
Boolean LastWasPeriod=false;
char c;
Buff = Alloc(BuffSize=50, char);
BN = 0;
while ( true )
{
c = InChar(f);
if ( c == '|' ) SkipComment;
if ( c == EOF || c == '\n' && LastWasPeriod )
{
/* The definition is complete. Add a period if it's
not there already and terminate the string */
if ( ! LastWasPeriod ) Append('.');
Append(0);
return;
}
if ( Space(c) )
{
Append(' ');
}
else
if ( c == '\\' )
{
/* Escaped character -- bypass any special meaning */
Append(InChar(f));
}
else
{
LastWasPeriod = ( c == '.' );
Append(c);
}
}
}
/*************************************************************************/
/* */
/* Append a character to Buff, resizing it if necessary */
/* */
/*************************************************************************/
void Append(char c)
/* ------ */
{
if ( c == ' ' && (! BN || Buff[BN-1] == ' ' ) ) return;
if ( BN >= BuffSize )
{
Realloc(Buff, BuffSize += 50, char);
}
Buff[BN++] = c;
}
/*************************************************************************/
/* */
/* Recursive descent parser with syntax error checking. */
/* The reverse polish is built up by calls to Dump() and DumpOp(), */
/* which also check for semantic validity. */
/* */
/* For possible error messages, each routine also keeps track of */
/* the beginning of the construct that it recognises (in Fi). */
/* */
/*************************************************************************/
Boolean Expression(void)
/* ---------- */
{
int Fi=BN;
if ( Buff[BN] == ' ' ) BN++;
if ( ! Conjunct() ) FailSyn("expression");
while ( Find("or") )
{
BN += 2;
if ( ! Conjunct() ) FailSyn("expression");
DumpOp(OP_OR, Fi);
}
return true;
}
Boolean Conjunct(void)
/* -------- */
{
int Fi=BN;
if ( ! SExpression() ) FailSyn("expression");
while ( Find("and") )
{
BN += 3;
if ( ! SExpression() ) FailSyn("expression");
DumpOp(OP_AND, Fi);
}
return true;
}
String RelOps[] = {">=", "<=", "!=", "<>", ">", "<", "=", (String) 0};
Boolean SExpression(void)
/* ----------- */
{
int o, Fi=BN;
if ( ! AExpression() ) FailSyn("expression");
if ( (o = FindOne(RelOps)) >= 0 )
{
BN += strlen(RelOps[o]);
if ( ! AExpression() ) FailSyn("expression");
DumpOp(( o == 0 ? OP_GE :
o == 1 ? OP_LE :
o == 4 ? OP_GT :
o == 5 ? OP_LT :
o == 2 || o == 3 ?
( TStack[TSN-1].Type == 'S' ? OP_SNE : OP_NE ) :
( TStack[TSN-1].Type == 'S' ? OP_SEQ : OP_EQ ) ), Fi);
}
return true;
}
String AddOps[] = {"+", "-", (String) 0};
Boolean AExpression(void)
/* ----------- */
{
int o, Fi=BN;
if ( Buff[BN] == ' ' ) BN++;
if ( (o = FindOne(AddOps)) >= 0 )
{
BN += 1;
}
if ( ! Term() ) FailSyn("expression");
if ( o == 1 ) DumpOp(OP_UMINUS, Fi);
while ( (o = FindOne(AddOps)) >= 0 )
{
BN += 1;
if ( ! Term() ) FailSyn("arithmetic expression");
DumpOp((char)(OP_PLUS + o), Fi);
}
return true;
}
String MultOps[] = {"*", "/", "%", (String) 0};
Boolean Term(void)
/* ---- */
{
int o, Fi=BN;
if ( ! Factor() ) FailSyn("expression");
while ( (o = FindOne(MultOps)) >= 0 )
{
BN += 1;
if ( ! Factor() ) FailSyn("arithmetic expression");
DumpOp((char)(OP_MULT + o), Fi);
}
return true;
}
Boolean Factor(void)
/* ---- */
{
int Fi=BN;
if ( ! Primary() ) FailSyn("value");
while ( Find("^") )
{
BN += 1;
if ( ! Primary() ) FailSyn("exponent");
DumpOp(OP_POW, Fi);
}
return true;
}
Boolean Primary(void)
/* ------- */
{
if ( Atom() )
{
return true;
}
else
if ( Find("(") )
{
BN++;
if ( ! Expression() ) FailSyn("expression in parentheses");
if ( ! Find(")") ) FailSyn("')'");
BN++;
return true;
}
else
{
FailSyn("attribute, value, or '('");
}
}
String Funcs[] = {"sin", "cos", "tan", "log", "exp", "int", (String) 0};
Boolean Atom(void)
/* ---- */
{
char *EndPtr, *Str, Date[11], Time[9];
int o, FirstBN, Fi=BN;
ContValue F;
Attribute Att;
if ( Buff[BN] == ' ' ) BN++;
if ( Buff[BN] == '"' )
{
FirstBN = ++BN;
while ( Buff[BN] != '"' )
{
if ( ! Buff[BN] ) FailSyn("closing '\"'");
BN++;
}
/* Make a copy of the string without double quotes */
Buff[BN] = '\00';
Str = strdup(Buff + FirstBN);
Buff[BN++] = '"';
Dump(OP_STR, 0, Str, Fi);
}
else
if ( (Att = FindAttName()) )
{
BN += strlen(AttName[Att]);
Dump(OP_ATT, 0, (String) (long) Att, Fi);
}
else
if ( isdigit(Buff[BN]) )
{
/* Check for date or time first */
if ( ( Buff[BN+4] == '/' && Buff[BN+7] == '/' ||
Buff[BN+4] == '-' && Buff[BN+7] == '-' )&&
isdigit(Buff[BN+1]) && isdigit(Buff[BN+2]) &&
isdigit(Buff[BN+3]) &&
isdigit(Buff[BN+5]) && isdigit(Buff[BN+6]) &&
isdigit(Buff[BN+8]) && isdigit(Buff[BN+9]) )
{
memcpy(Date, Buff+BN, 10);
Date[10] = '\00';
if ( (F = DateToDay(Date)) == 0 )
{
Error(BADDEF1, Date, "date");
}
BN += 10;
}
else
if ( Buff[BN+2] == ':' && Buff[BN+5] == ':' &&
isdigit(Buff[BN+1]) &&
isdigit(Buff[BN+3]) && isdigit(Buff[BN+4]) &&
isdigit(Buff[BN+6]) && isdigit(Buff[BN+7]) )
{
memcpy(Time, Buff+BN, 8);
Time[8] = '\00';
if ( (F = TimeToSecs(Time)) == 0 )
{
Error(BADDEF1, Time, "time");
}
BN += 8;
}
else
{
F = strtod(Buff+BN, &EndPtr);
/* Check for period after integer */
if ( EndPtr > Buff+BN+1 && *(EndPtr-1) == '.' )
{
EndPtr--;
}
BN = EndPtr - Buff;
}
Dump(OP_NUM, F, Nil, Fi);
}
else
if ( (o = FindOne(Funcs)) >= 0 )
{
BN += 3;
if ( ! Find("(") ) FailSyn("'(' after function name");
BN++;
if ( ! Expression() ) FailSyn("expression");
if ( ! Find(")") ) FailSyn("')' after function argument");
BN++;
DumpOp((char)(OP_SIN + o), Fi);
}
else
if ( Buff[BN] == '?' )
{
BN++;
if ( TStack[TSN-1].Type == 'N' )
{
Dump(OP_NUM, _UNK.cval, Nil, Fi);
}
else
{
Dump(OP_STR, 0, Nil, Fi);
}
}
else
if ( ! memcmp(Buff+BN, "N/A", 3) )
{
BN += 3;
if ( TStack[TSN-1].Type == 'N' )
{
Dump(OP_NUM, _NA.cval, Nil, Fi);
}
else
{
Dump(OP_STR, 0, strdup("N/A"), Fi);
}
}
else
{
return false;
}
return true;
}
/*************************************************************************/
/* */
/* Skip spaces and check for specific string */
/* */
/*************************************************************************/
Boolean Find(String S)
/* ---- */
{
if ( Buff[BN] == ' ' ) BN++;
return ( ! Buff[BN] ? false : ! memcmp(Buff+BN, S, strlen(S)) );
}
/*************************************************************************/
/* */
/* Find one of a zero-terminated list of alternatives */
/* */
/*************************************************************************/
int FindOne(String *Alt)
/* ------- */
{
int a;
for ( a = 0 ; Alt[a] ; a++ )
{
if ( Find(Alt[a]) ) return a;
}
return -1;
}
/*************************************************************************/
/* */
/* Find an attribute name */
/* */
/*************************************************************************/
Attribute FindAttName(void)
/* ----------- */
{
Attribute Att, LongestAtt=0;
ForEach(Att, 1, MaxAtt-1)
{
if ( ! Exclude(Att) && Find(AttName[Att]) )
{
if ( ! LongestAtt ||
strlen(AttName[Att]) > strlen(AttName[LongestAtt]) )
{
LongestAtt = Att;
}
}
}
return LongestAtt;
}
/*************************************************************************/
/* */
/* Error message routines. Syntax errors come from the */
/* recursive descent parser, semantics errors from the routines */
/* that build up the equivalent polish */
/* */
/*************************************************************************/
void DefSyntaxError(String Msg)
/* -------------- */
{
String RestOfText;
int i=10;
if ( ! PreviousError )
{
RestOfText = Buff + BN;
/* Abbreviate text if longer than 12 characters */
if ( CharWidth(RestOfText) > 12 )
{
#ifdef UTF8
/* Find beginning of UTF-8 character */
for ( ; (RestOfText[i] & 0x80) ; i++)
;
#endif
RestOfText[i] = RestOfText[i+1] = '.';
}
Error(BADDEF1, RestOfText, Msg);
PreviousError = true;
}
}
void DefSemanticsError(int Fi, String Msg, int OpCode)
/* ----------------- */
{
char Exp[1000], XMsg[1000], Op[1000];
if ( ! PreviousError )
{
/* Abbreviate the input if necessary */
if ( BN - Fi > 23 )
{
sprintf(Exp, "%.10s...%.10s", Buff+Fi, Buff+BN-10);
}
else
{
sprintf(Exp, "%.*s", BN - Fi, Buff+Fi);
}
switch ( OpCode )
{
case OP_AND: sprintf(Op, "%s", "and"); break;
case OP_OR: sprintf(Op, "%s", "or"); break;
case OP_SEQ:
case OP_EQ: sprintf(Op, "%s", "="); break;
case OP_SNE:
case OP_NE: sprintf(Op, "%s", "<>"); break;
case OP_GT: sprintf(Op, "%s", ">"); break;
case OP_GE: sprintf(Op, "%s", ">="); break;
case OP_LT: sprintf(Op, "%s", "<"); break;
case OP_LE: sprintf(Op, "%s", "<="); break;
case OP_PLUS: sprintf(Op, "%s", "+"); break;
case OP_MINUS: sprintf(Op, "%s", "-"); break;
case OP_UMINUS: sprintf(Op, "%s", "unary -"); break;
case OP_MULT: sprintf(Op, "%s", "*"); break;
case OP_DIV: sprintf(Op, "%s", "/"); break;
case OP_MOD: sprintf(Op, "%s", "%"); break;
case OP_POW: sprintf(Op, "%s", "^"); break;
case OP_SIN: sprintf(Op, "%s", "sin"); break;
case OP_COS: sprintf(Op, "%s", "cos"); break;
case OP_TAN: sprintf(Op, "%s", "tan"); break;
case OP_LOG: sprintf(Op, "%s", "log"); break;
case OP_EXP: sprintf(Op, "%s", "exp"); break;
case OP_INT: sprintf(Op, "%s", "int");
}
sprintf(XMsg, "%s with '%s'", Msg, Op);
Error(BADDEF2, Exp, XMsg);
PreviousError = true;
}
}
/*************************************************************************/
/* */
/* Reverse polish routines. These use a model of the stack */
/* during expression evaluation to detect type conflicts etc */
/* */
/*************************************************************************/
void Dump(char OpCode, ContValue F, String S, int Fi)
/* ---- */
{
if ( Buff[Fi] == ' ' ) Fi++;
if ( ! UpdateTStack(OpCode, F, S, Fi) ) return;
/* Make sure enough room for this element */
if ( DN >= DefSize-1 )
{
Realloc(AttDef[MaxAtt], DefSize += 100, DefElt);
}
DefOp(AttDef[MaxAtt][DN]) = OpCode;
if ( OpCode == OP_ATT || OpCode == OP_STR )
{
DefSVal(AttDef[MaxAtt][DN]) = S;
}
else
{
DefNVal(AttDef[MaxAtt][DN]) = F;
}
DN++;
}
void DumpOp(char OpCode, int Fi)
/* ------ */
{
Dump(OpCode, 0, Nil, Fi);
}
Boolean UpdateTStack(char OpCode, ContValue F, String S, int Fi)
/* ------------ */
{
if ( TSN >= TStackSize )
{
Realloc(TStack, TStackSize += 50, EltRec);
}
switch ( OpCode )
{
case OP_ATT:
TStack[TSN].Type = ( Continuous((long) S) ? 'N' : 'S' );
break;
case OP_NUM:
TStack[TSN].Type = 'N';
break;
case OP_STR:
TStack[TSN].Type = 'S';
break;
case OP_AND:
case OP_OR:
if ( TStack[TSN-2].Type != 'B' || TStack[TSN-1].Type != 'B' )
{
FailSem("non-logical value");
}
TSN -= 2;
break;
case OP_EQ:
case OP_NE:
if ( TStack[TSN-2].Type != TStack[TSN-1].Type )
{
FailSem("incompatible values");
}
TSN -= 2;
TStack[TSN].Type = 'B';
break;
case OP_GT:
case OP_GE:
case OP_LT:
case OP_LE:
if ( TStack[TSN-2].Type != 'N' || TStack[TSN-1].Type != 'N' )
{
FailSem("non-arithmetic value");
}
TSN -= 2;
TStack[TSN].Type = 'B';
break;
case OP_SEQ:
case OP_SNE:
if ( TStack[TSN-2].Type != 'S' || TStack[TSN-1].Type != 'S' )
{
FailSem("incompatible values");
}
TSN -= 2;
TStack[TSN].Type = 'B';
break;
case OP_PLUS:
case OP_MINUS:
case OP_MULT:
case OP_DIV:
case OP_MOD:
case OP_POW:
if ( TStack[TSN-2].Type != 'N' || TStack[TSN-1].Type != 'N' )
{
FailSem("non-arithmetic value");
}
TSN -= 2;
break;
case OP_UMINUS:
if ( TStack[TSN-1].Type != 'N' )
{
FailSem("non-arithmetic value");
}
TSN--;
break;
case OP_SIN:
case OP_COS:
case OP_TAN:
case OP_LOG:
case OP_EXP:
case OP_INT:
if ( TStack[TSN-1].Type != 'N' )
{
FailSem("non-arithmetic argument");
}
TSN--;
}
TStack[TSN].Fi = Fi;
TStack[TSN].Li = BN-1;
TSN++;
return true;
}
/*************************************************************************/
/* */
/* Evaluate an implicit attribute for a case */
/* */
/*************************************************************************/
#define CUnknownVal(AV) (AV.cval==_UNK.cval)
#define DUnknownVal(AV) (! AV.dval)
#define DUNA(a) (DUnknownVal(XStack[a]) || NotApplicVal(XStack[a]))
#define CUNA(a) (CUnknownVal(XStack[a]) || NotApplicVal(XStack[a]))
#define C1(x) (CUNA(XSN-1) ? _UNK.cval : (x))
#define C2(x) (CUNA(XSN-1) || CUNA(XSN-2) ? _UNK.cval : (x))
#define CD2(x) (CUNA(XSN-1) || CUNA(XSN-2) ? 0 : (x))
#define D2(x) (DUNA(XSN-1) || DUNA(XSN-2) ? 0 : (x))
AttValue EvaluateDef(Definition D, DataRec Case)
/* ----------- */
{
XStackElt XStack[100]; /* allows 100-level nesting */
int XSN=0, DN, bv1, bv2, Mult;
double cv1, cv2;
String sv1, sv2;
Attribute Att;
DefElt DElt;
AttValue ReturnVal;
for ( DN = 0 ; ; DN++)
{
switch ( DefOp((DElt = D[DN])) )
{
case OP_ATT:
Att = (long) DefSVal(DElt);
if ( Continuous(Att) )
{
XStack[XSN++].cval = CVal(Case, Att);
}
else
{
XStack[XSN++].sval =
( Unknown(Case, Att) && ! NotApplic(Case, Att) ? 0 :
AttValName[Att][XDVal(Case, Att)] );
}
break;
case OP_NUM:
XStack[XSN++].cval = DefNVal(DElt);
break;
case OP_STR:
XStack[XSN++].sval = DefSVal(DElt);
break;
case OP_AND:
bv1 = XStack[XSN-2].dval;
bv2 = XStack[XSN-1].dval;
XStack[XSN-2].dval = ( bv1 == 3 || bv2 == 3 ? 3 :
D2(bv1 == 2 && bv2 == 2 ? 2 : 3) );
XSN--;
break;
case OP_OR:
bv1 = XStack[XSN-2].dval;
bv2 = XStack[XSN-1].dval;
XStack[XSN-2].dval = ( bv1 == 2 || bv2 == 2 ? 2 :
D2(bv1 == 2 || bv2 == 2 ? 2 : 3) );
XSN--;
break;
case OP_EQ:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = ( cv1 == cv2 ? 2 : 3 );
XSN--;
break;
case OP_NE:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = ( cv1 != cv2 ? 2 : 3 );
XSN--;
break;
case OP_GT:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = CD2(cv1 > cv2 ? 2 : 3);
XSN--;
break;
case OP_GE:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = CD2(cv1 >= cv2 ? 2 : 3);
XSN--;
break;
case OP_LT:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = CD2(cv1 < cv2 ? 2 : 3);
XSN--;
break;
case OP_LE:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].dval = CD2(cv1 <= cv2 ? 2 : 3);
XSN--;
break;
case OP_SEQ:
sv1 = XStack[XSN-2].sval;
sv2 = XStack[XSN-1].sval;
XStack[XSN-2].dval =
( ! sv1 && ! sv2 ? 2 :
! sv1 || ! sv2 ? 3 :
! strcmp(sv1, sv2) ? 2 : 3 );
XSN--;
break;
case OP_SNE:
sv1 = XStack[XSN-2].sval;
sv2 = XStack[XSN-1].sval;
XStack[XSN-2].dval =
( ! sv1 && ! sv2 ? 3 :
! sv1 || ! sv2 ? 2 :
strcmp(sv1, sv2) ? 2 : 3 );
XSN--;
break;
case OP_PLUS:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].cval = C2(cv1 + cv2);
XSN--;
break;
case OP_MINUS:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].cval = C2(cv1 - cv2);
XSN--;
break;
case OP_MULT:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].cval = C2(cv1 * cv2);
XSN--;
break;
case OP_DIV:
/* Note: have to set precision of result */
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
if ( ! cv2 ||
CUnknownVal(XStack[XSN-2]) ||
CUnknownVal(XStack[XSN-1]) ||
NotApplicVal(XStack[XSN-2]) ||
NotApplicVal(XStack[XSN-1]) )
{
XStack[XSN-2].cval = _UNK.cval;
}
else
{
Mult = Denominator(cv1);
cv1 = cv1 / cv2;
while ( fabs(cv2) > 1 )
{
Mult *= 10;
cv2 /= 10;
}
XStack[XSN-2].cval = rint(cv1 * Mult) / Mult;
}
XSN--;
break;
case OP_MOD:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].cval = C2(fmod(cv1, cv2));
XSN--;
break;
case OP_POW:
cv1 = XStack[XSN-2].cval;
cv2 = XStack[XSN-1].cval;
XStack[XSN-2].cval =
( CUNA(XSN-1) || CUNA(XSN-2) ||
( cv1 < 0 && ceil(cv2) != cv2 ) ? _UNK.cval :
pow(cv1, cv2) );
XSN--;
break;
case OP_UMINUS:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(-cv1);
break;
case OP_SIN:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(sin(cv1));
break;
case OP_COS:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(cos(cv1));
break;
case OP_TAN:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(tan(cv1));
break;
case OP_LOG:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval =
( CUNA(XSN-1) || cv1 <= 0 ? _UNK.cval : log(cv1) );
break;
case OP_EXP:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(exp(cv1));
break;
case OP_INT:
cv1 = XStack[XSN-1].cval;
XStack[XSN-1].cval = C1(rint(cv1));
break;
case OP_END:
ReturnVal.cval = XStack[0].cval; /* cval >= dval bytes */
return ReturnVal;
}
}
}
instance.c 0000664 0000764 0000764 00000050307 11501534415 012744 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define DPREC 16 /* distance precision */
#define AdjustedValue(i,b) (Cttee?(Class(Instance[i])+b-RSPredVal[i]):\
Class(Instance[i]))
#define CVDiff(c,cv,a) (fabs(CVal(c,a)-(cv))/(5*AttSD[a]))
#define Select(t) (UseAll? (t) :\
(MaxInstance+1)*((2*(t)+1)/(2.0*Try)))
int MinN; /* minimum close neighbors */
int Try; /* sample size for estimates */
Boolean UseAll, /* true if no sampling */
SetNN, /* true if NN set automatically */
*Tested=Nil; /* for BuildIndex */
CaseCount
*ValFreq=Nil; /* ditto */
AttValue *KDBlock; /* copy of instances in KDTree order to
improve caching */
/*************************************************************************/
/* */
/* Prepare for instance-based prediction */
/* */
/* The modified K-D tree algorithm adds to each instance the */
/* distances to two "reference" points. The first is an artifical */
/* point intended to play the role of an origin: its value for an */
/* attribute should be lower than most of the instances' values. */
/* The second reference point is the instance that is furthest */
/* from the first reference point. */
/* */
/*************************************************************************/
void InitialiseInstances(RRuleSet *Cttee)
/* ------------------- */
{
CaseNo i, FarInstance;
Attribute Att;
double Dist;
/* Use the data items as instances. The vector Instance[] must
be allocated separately since the indexing re-orders the items */
Instance = Alloc(MaxCase+1, DataRec);
MaxInstance = MaxCase;
ForEach(i, 0, MaxCase)
{
Instance[i] = Case[i];
}
Tested = AllocZero(MaxAtt+1, unsigned char);
ValFreq = Alloc(MaxDiscrVal+1, CaseCount);
GNNEnv.AttMinD = Alloc(MaxAtt+1, float);
/* Construct the first reference point */
Ref[0] = Alloc(MaxAtt+1, AttValue);
ForEach(Att, 1, MaxAtt)
{
if ( Continuous(Att) )
{
CVal(Ref[0], Att) = AttMean[Att] - 2.5 * AttSD[Att];
}
else
{
DVal(Ref[0], Att) = 2;
}
}
/* Compute distances and find an instance as far as possible from
the first reference point */
FarInstance = 0;
ForEach(i, 0, MaxInstance)
{
DRef1(Instance[i]) = Dist = Distance(Instance[i], Ref[0], 1E38);
if ( Dist > DRef1(Instance[FarInstance]) ) FarInstance = i;
}
Ref[1] = Alloc(MaxAtt+1, AttValue);
memcpy(Ref[1], Instance[FarInstance], (MaxAtt+1) * sizeof(AttValue));
/* Now compute distances to the second reference point */
ForEach(i, 0, MaxInstance)
{
DRef2(Instance[i]) = Dist = Distance(Instance[i], Ref[1], 1E38);
}
NotifyStage(INDEXINSTANCES);
Progress(-1.0);
KDTree = BuildIndex(0, MaxCase);
Free(Tested); Tested = Nil;
Free(ValFreq); ValFreq = Nil;
/* Tabulate values predicted by ruleset. Cannot use
FindPredictedValues() because that would use instances! */
RSPredVal = Alloc(MaxCase+1, float);
ForEach(i, 0, MaxCase)
{
RSPredVal[i] = PredictValue(Cttee, Instance[i]);
}
/* Set parameters for nearest neighbors if not read from model file */
Try = Min(MaxInstance+1, 1000);
UseAll = ( Try == MaxInstance+1 );
if ( MAXD < 0 )
{
SetParameters(Cttee);
}
else
{
SetNN = false;
}
MinN = (NN + 1) / 2;
GNNEnv.WorstBest = GNNEnv.BestD + NN-1;
}
/*************************************************************************/
/* */
/* Set "close" distance MAXD to average distance. */
/* If required, find the best NN value between 1 and NNMAX */
/* */
/*************************************************************************/
void SetParameters(RRuleSet *Cttee)
/* ------------- */
{
double TotErr[NNMAX+1], Sum=0;
float Estimate, RealClass;
CaseNo i, j;
int t, BestNN=1;
/* Set MAXD to the average distance between instances */
GNNEnv.WorstBest = GNNEnv.BestD; /* use single nearest neighbor */
ForEach(t, 0, Try-1)
{
i = Select(t);
while ( (j = KRandom() * (MaxInstance+1)) == i )
;
Sum += Distance(Instance[j], Instance[i], 1E10);
}
MAXD = rint(DPREC * Sum / Try) / DPREC;
/* Set NN if required */
if ( NN )
{
SetNN = false;
}
else
{
SetNN = true;
NotifyStage(SETNEIGHBORS);
Progress( -Try );
/* Estimate error with different values of NN */
ForEach(NN, 1, NNMAX)
{
TotErr[NN] = 0;
}
GNNEnv.WorstBest = GNNEnv.BestD + NNMAX-1;
ForEach(t, 0, Try-1)
{
i = Select(t);
RealClass = Class(Instance[i]);
FindNearestNeighbors(Instance[i]);
ForEach(NN, 1, NNMAX)
{
MinN = (NN + 1) / 2;
Estimate = AverageNeighbors(Cttee, Instance[i]);
TotErr[NN] += fabs(RealClass - Estimate);
}
Progress(1.0);
}
/* Select the best number of neighbors */
ForEach(NN, 2, NNMAX)
{
if ( TotErr[NN] < TotErr[BestNN] )
{
BestNN = NN;
}
}
ForEach(NN, 1, BestNN-1)
{
if ( TotErr[NN] < 1.01 * TotErr[BestNN] )
{
BestNN = NN;
break;
}
}
NN = BestNN;
fprintf(Of, T_SettingNNeighbors, NN);
}
}
/*************************************************************************/
/* */
/* If allowed to choose, decide whether to use instances based */
/* on estimated errors of rule-based and composite models */
/* */
/*************************************************************************/
void CheckForms(RRuleSet *Cttee)
/* ---------- */
{
CaseNo i;
int t;
double RSErr=0, IErr=0, RSIErr=0, AvRules=0;
ContValue RealClass;
NotifyStage(ASSESSCOMPOSITE);
Progress( -Try );
ForEach(t, 0, Try-1)
{
i = Select(t);
RealClass = Class(Case[i]);
FindNearestNeighbors(Case[i]);
IErr += fabs(RealClass - AverageNeighbors(Nil, Case[i]));
RSErr += fabs(RealClass - PredictValue(Cttee, Case[i]));
RSIErr += fabs(RealClass - AverageNeighbors(Cttee, Case[i]));
Progress(1.0);
}
/* Find average number of rules in cttee */
ForEach(i, 0, MEMBERS-1)
{
AvRules += Cttee[i]->SNRules;
}
AvRules /= MEMBERS;
if ( ! SetNN ) fprintf(Of, "\n\n");
Verbosity(1,
fprintf(Of, "\n\n %s |error| on training data:\n",
( UseAll ? "Average" : "Sampled" ));
fprintf(Of, "\tinstances:\t\t%.2f\n", IErr / Try);
fprintf(Of, "\trules:\t\t\t%.2f\n", RSErr / Try);
fprintf(Of, "\trules + instances:\t%.2f\n", RSIErr / Try))
if ( EstimateErr(RSErr, MaxCase+1, 2 * AvRules) <
EstimateErr(RSIErr, MaxCase+1, AvRules) )
{
USEINSTANCES = false;
fprintf(Of, T_SuggestRules);
FreeInstances();
}
else
{
fprintf(Of, T_SuggestComposite);
}
}
/*************************************************************************/
/* */
/* Make a copy of instances in KDTree order to improve caching */
/* */
/*************************************************************************/
void CopyInstances(void)
/* ------------- */
{
AttValue *KDBlockP;
CaseNo i;
KDBlockP = KDBlock = Alloc((MaxInstance+1) * (MaxAtt+3), AttValue);
ForEach(i, 0, MaxInstance)
{
memcpy(KDBlockP, Instance[i], (MaxAtt+3) * sizeof(AttValue));
Instance[i] = KDBlockP;
KDBlockP += MaxAtt+3;
}
}
/*************************************************************************/
/* */
/* Use nearest neighbors to estimate target for Case. */
/* If Cttee is not null, values of neighbors are adjusted by */
/* rulesets before being combined. */
/* */
/*************************************************************************/
float NNEstimate(RRuleSet *Cttee, DataRec Case)
/* ---------- */
{
FindNearestNeighbors(Case);
return AverageNeighbors(Cttee, Case);
}
/*************************************************************************/
/* */
/* Determine the distance between two cases. */
/* Stop if the distance becomes too great to be relevant. */
/* */
/* Cubist uses Manhattan distance between instances (i.e. sum */
/* of the attribute differences). These are defined as: */
/* * discrete att: 0 (if same) or 2 / number of discrete values */
/* * continuous att: difference / 5 SD */
/* */
/*************************************************************************/
float Distance(DataRec Case1, DataRec Case2, float Thresh)
/* -------- */
{
Attribute Att;
double DTot, Diff;
for ( Att = 1, DTot = 0 ; DTot < Thresh && Att <= MaxAtt ; Att++ )
{
if ( Skip(Att) || Att == ClassAtt ) continue;
if ( NotApplic(Case2, Att) != NotApplic(Case1, Att) )
{
DTot += 1.0;
}
else
if ( Continuous(Att) )
{
Diff = CVDiff(Case2, CVal(Case1, Att), Att);
DTot += Min(1.0, Diff);
}
else
if ( Ordered(Att) )
{
DTot += fabs(DVal(Case2, Att) - DVal(Case1, Att)) /
(MaxAttVal[Att] - 1);
}
else
if ( DVal(Case2, Att) != DVal(Case1, Att) )
{
DTot += 2.0 / (MaxAttVal[Att] - 1);
}
}
return DTot;
}
/*************************************************************************/
/* */
/* Check whether a saved instance should be one of the neighbors. */
/* Distances are rounded to precision 1/DPREC */
/* */
/*************************************************************************/
void CheckDistance(DataRec Case, CaseNo Saved)
/* ------------- */
{
int d, dd;
float Dist;
if ( Instance[Saved] == Case ) return;
Dist = rint(DPREC *
Distance(Case, Instance[Saved], *GNNEnv.WorstBest + 0.55 / DPREC))
/ DPREC;
if ( Dist <= *GNNEnv.WorstBest )
{
for ( d = 0 ; d < MAXN && GNNEnv.BestD[d] < Dist ; d++ )
;
if ( d < MAXN )
{
for ( dd = MAXN-1 ; dd > d ; dd-- )
{
GNNEnv.BestD[dd] = GNNEnv.BestD[dd-1];
GNNEnv.BestI[dd] = GNNEnv.BestI[dd-1];
}
GNNEnv.BestD[d] = Dist;
GNNEnv.BestI[d] = Saved;
}
}
}
/*************************************************************************/
/* */
/* Find the NN nearest saved instances to a given case. */
/* KD-trees are used to locate neighbors quickly. */
/* */
/*************************************************************************/
void FindNearestNeighbors(DataRec Case)
/* -------------------- */
{
int d;
Attribute Att;
/* Clear best distances and attribute minimum distances */
ForEach(d, 0, MAXN-1)
{
GNNEnv.BestD[d] = MAXD;
GNNEnv.BestI[d] = -1;
}
ForEach(Att, 1, MaxAtt)
{
GNNEnv.AttMinD[Att] = 0;
}
DRef1(Case) = Distance(Case, Ref[0], 1E38);
DRef2(Case) = Distance(Case, Ref[1], 1E38);
ScanIndex(Case, KDTree, 0.0);
}
/*************************************************************************/
/* */
/* Find weighted average value of selected neighbors. */
/* [Weight = 1 / (Manhattan distance + 0.5).] */
/* Have to be careful with ties. */
/* */
/*************************************************************************/
float AverageNeighbors(RRuleSet *Cttee, DataRec Case)
/* ---------------- */
{
int d=0, Count=0, Same, Last;
double Est, BaseEst, Wt, SameSum, SameWt, TotSum=0, TotWt=0;
BaseEst = ( Cttee ? PredictValue(Cttee, Case) : GlobalMean );
/* Check the number of neighbors actually found */
for ( Last = MAXN-1 ; Last >= 0 && GNNEnv.BestI[Last] < 0 ; Last-- )
;
if ( Last+1 < MinN )
{
return BaseEst;
}
/* Extract groups of neighbors with the same values of BestD */
while ( d <= Last && Count < NN && Count < MaxInstance )
{
SameSum = SameWt = Same = 0;
Wt = 1 / (GNNEnv.BestD[d] + 0.5);
do
{
Est = AdjustedValue(GNNEnv.BestI[d], BaseEst);
if ( Est > Ceiling ) Est = Ceiling;
else
if ( Est < Floor ) Est = Floor;
SameSum += Wt * Est;
SameWt += Wt;
Same++;
d++;
}
while ( d <= Last && GNNEnv.BestD[d] == GNNEnv.BestD[d-1] );
if ( Count + Same > NN )
{
Wt = (NN - Count) / (float) Same;
TotSum += Wt * SameSum;
TotWt += Wt * SameWt;
Count = NN;
}
else
{
TotSum += SameSum;
TotWt += SameWt;
Count += Same;
}
}
Est = TotSum / TotWt;
return ( Est < Floor ? Floor : Est > Ceiling ? Ceiling : Est );
}
/*************************************************************************/
/* */
/* The following routines are concerned with indexing the */
/* instances in a KD-tree, and using the tree to locate nearest */
/* neighbors without having to examine each saved instance. */
/* Note that BuildIndex() can create branches that have no cases */
/* associated with them -- this is ok since it speeds up */
/* ScanIndex() by increasing MinD. */
/* */
/*************************************************************************/
Index BuildIndex(CaseNo Fp, CaseNo Lp)
/* ---------- */
{
Index Node;
DiscrValue v, vv;
CaseNo i, Xp, Kp;
CaseCount Cases;
double Mean, BestMean, ExpDist, BestExpDist=0, ProbNA;
float Dist, MinDRef[2], MaxDRef[2];
Attribute Att, BestAtt=0;
if ( Lp < Fp ) return Nil;
Node = AllocZero(1, IndexRec);
if ( Lp > Fp )
{
MinDRef[0] = MaxDRef[0] = DRef1(Instance[Fp]);
MinDRef[1] = MaxDRef[1] = DRef2(Instance[Fp]);
ForEach(i, Fp+1, Lp)
{
if ( (Dist = DRef1(Instance[i])) < MinDRef[0] )
{
MinDRef[0] = Dist;
}
else
if ( Dist > MaxDRef[0] )
{
MaxDRef[0] = Dist;
}
if ( (Dist = DRef2(Instance[i])) < MinDRef[1] )
{
MinDRef[1] = Dist;
}
else
if ( Dist > MaxDRef[1] )
{
MaxDRef[1] = Dist;
}
}
Node->MinDRef[0] = MinDRef[0];
Node->MaxDRef[0] = MaxDRef[0];
Node->MinDRef[1] = MinDRef[1];
Node->MaxDRef[1] = MaxDRef[1];
/* Find the attribute with the greatest expected difference.
Distances to the reference points are candidates */
ForEach(Att, 1, MaxAtt)
{
if ( Skip(Att) || Att == ClassAtt || Tested[Att] && Discrete(Att) )
{
continue;
}
/* Separate all N/A values (real attributes only) */
Xp = Fp;
ForEach(i, Fp, Lp)
{
if ( NotApplic(Instance[i], Att) )
{
SwapInstance(i, Xp++);
}
}
ProbNA = (Xp - Fp) / (Lp - Fp + 1.0);
if ( ! (Cases = Lp-Xp+1) ) continue;
ExpDist = Mean = 0;
if ( Continuous(Att) )
{
/* Expected distance is average difference from mean */
ForEach(i, Xp, Lp)
{
Mean += CVal(Instance[i], Att);
}
Mean /= Cases;
ForEach(i, Xp, Lp)
{
ExpDist += CVDiff(Instance[i], Mean, Att);
}
ExpDist /= Cases;
}
else
{
/* Expected distance is computed from pairwise differences
of values */
ForEach(v, 2, MaxAttVal[Att])
{
ValFreq[v] = 0;
}
ForEach(i, Xp, Lp)
{
ValFreq[DVal(Instance[i], Att)] ++;
}
if ( Ordered(Att) )
{
ForEach(v, 2, MaxAttVal[Att])
{
ForEach(vv, 2, MaxAttVal[Att])
{
ExpDist += ValFreq[v] * ValFreq[vv] * fabs(vv - v);
}
}
}
else
{
ForEach(v, 2, MaxAttVal[Att])
{
ExpDist += ValFreq[v] * (Cases - ValFreq[v]) * 2.0;
}
}
ExpDist /= (MaxAttVal[Att] - 1) * Cases * Cases;
}
/* Final expected distance =
(prob one or other N/A) * 1 +
(prob both known) * (expected distance if known) */
ExpDist = 2 * ProbNA * (1 - ProbNA) +
(1 - ProbNA) * (1 - ProbNA) * ExpDist;
if ( ExpDist > BestExpDist )
{
BestExpDist = ExpDist;
BestAtt = Att;
BestMean = Mean;
}
}
}
/* Check whether leaf or sub-index */
if ( ! BestAtt )
{
Node->Tested = 0;
Node->IFp = Fp;
Node->ILp = Lp;
}
else
if ( Discrete(BestAtt) )
{
Node->Tested = BestAtt;
Node->SubIndex = Alloc(MaxAttVal[BestAtt]+1, Index);
Tested[BestAtt] = true;
/* Sort instances by attribute value */
Kp = Fp;
ForEach(v, 1, MaxAttVal[BestAtt])
{
ForEach(Xp, Kp, Lp)
{
if ( DVal(Instance[Xp], BestAtt) == v )
{
SwapInstance(Xp, Kp++);
}
}
Node->SubIndex[v] = BuildIndex(Fp, Kp-1);
Fp = Kp;
}
Tested[BestAtt] = false;
}
else
{
Node->Tested = BestAtt;
Node->Cut = BestMean;
/* There are three branches for continuous attributes:
N/A / value <= mean / value > mean */
Xp = Fp;
ForEach(i, Fp, Lp)
{
if ( NotApplic(Instance[i], BestAtt) )
{
SwapInstance(i, Xp++);
}
}
Kp = Xp;
ForEach(i, Xp, Lp)
{
if ( CVal(Instance[i], BestAtt) <= BestMean )
{
SwapInstance(i, Kp++);
}
}
/* Safety check: return leaf if not a sensible split */
if ( Xp == Lp+1 ||
Xp == Fp && Kp == Lp+1 ||
Kp == Fp )
{
Node->Tested = 0;
Node->IFp = Fp;
Node->ILp = Lp;
}
else
{
Node->SubIndex = Alloc(4, Index);
Node->SubIndex[1] = BuildIndex(Fp, Xp-1);
Node->SubIndex[2] = BuildIndex(Xp, Kp-1);
Node->SubIndex[3] = BuildIndex(Kp, Lp);
}
}
return Node;
}
void SwapInstance(CaseNo A, CaseNo B)
/* ------------ */
{
DataRec Hold;
Hold = Instance[A];
Instance[A] = Instance[B];
Instance[B] = Hold;
}
void ScanIndex(DataRec Case, Index Node, float MinD)
/* --------- */
{
CaseNo Xp;
DiscrValue Forks, First, v;
float NewMinD, SaveAttMinD;
Attribute Att;
if ( Node == Nil ) return;
if ( ! (Att = Node->Tested) )
{
ForEach(Xp, Node->IFp, Node->ILp)
{
CheckDistance(Case, Xp);
}
}
else
if ( Max(Node->MinDRef[0] - DRef1(Case), DRef1(Case) - Node->MaxDRef[0]) <=
*GNNEnv.WorstBest + 0.5 / DPREC &&
Max(Node->MinDRef[1] - DRef2(Case), DRef2(Case) - Node->MaxDRef[1]) <=
*GNNEnv.WorstBest + 0.5 / DPREC )
{
if ( Discrete(Att) )
{
First = DVal(Case, Att);
Forks = MaxAttVal[Att];
}
else
{
First = ( NotApplic(Case, Att) ? 1 :
CVal(Case, Att) <= Node->Cut ? 2 : 3 );
Forks = 3;
}
/* Try best sub-index first, then other sub-indices so long
as can improve on current best neighbors */
if ( First <= Forks )
{
ScanIndex(Case, Node->SubIndex[First], MinD);
}
SaveAttMinD = GNNEnv.AttMinD[Att];
ForEach(v, 1, Forks)
{
if ( v == First || ! Node->SubIndex[v] ) continue;
GNNEnv.AttMinD[Att] =
( v == 1 || First == 1 ? 1.0 :
Continuous(Att) ? CVDiff(Case, Node->Cut, Att) :
Ordered(Att) ?
fabs(v - First) / (MaxAttVal[Att] - 1) :
2.0 / (MaxAttVal[Att] - 1) );
NewMinD = MinD + GNNEnv.AttMinD[Att] - SaveAttMinD;
if ( NewMinD <= *GNNEnv.WorstBest + 0.5 / DPREC )
{
ScanIndex(Case, Node->SubIndex[v], NewMinD);
}
}
GNNEnv.AttMinD[Att] = SaveAttMinD;
}
}
void FreeIndex(Index Node)
/* --------- */
{
DiscrValue v, Forks;
Attribute Att;
if ( Node == Nil ) return;
if ( (Att = Node->Tested) )
{
Forks = ( Discrete(Att) ? MaxAttVal[Att] : 3 );
ForEach(v, 1, Forks)
{
FreeIndex(Node->SubIndex[v]);
}
Free(Node->SubIndex);
}
Free(Node);
}
void FreeInstances(void)
/* ------------- */
{
if ( Instance )
{
Free(Instance); Instance = Nil;
Free(KDBlock); KDBlock = Nil;
}
FreeUnlessNil(GNNEnv.AttMinD); GNNEnv.AttMinD = Nil;
FreeUnlessNil(RSPredVal); RSPredVal = Nil;
if ( KDTree )
{
FreeUnlessNil(Ref[0]); Ref[0] = Nil;
FreeUnlessNil(Ref[1]); Ref[1] = Nil;
FreeIndex(KDTree); KDTree = Nil;
}
}
modelfiles.c 0000664 0000764 0000764 00000041737 11501534415 013272 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Routines for saving and reading model files */
/* ------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
int Entry;
char* Prop[]={"null",
"id",
"att",
"elts",
"prec",
"globalmean",
"floor",
"ceiling",
"sample",
"init",
"mean",
"sd",
"mode",
"entries",
"rules",
"cover",
"loval",
"hival",
"extrap",
"insts",
"nn",
"maxd",
"esterr",
"conds",
"type",
"cut",
"result",
"val",
"coeff",
"max",
"min",
"redn"
};
char PropName[20],
*PropVal=Nil,
*Unquoted;
int PropValSize=0;
#define PROPS 31
#define IDP 1
#define ATTP 2
#define ELTSP 3
#define PRECP 4
#define GLOBALMEANP 5
#define FLOORP 6
#define CEILINGP 7
#define SAMPLEP 8
#define INITP 9
#define MEANP 10
#define SDP 11
#define MODEP 12
#define ENTRIESP 13
#define RULESP 14
#define COVERP 15
#define LOVALP 16
#define HIVALP 17
#define EXTRAPP 18
#define INSTSP 19
#define NNP 20
#define MAXDP 21
#define ESTERRP 22
#define CONDSP 23
#define TYPEP 24
#define CUTP 25
#define RESULTP 26
#define VALP 27
#define COEFFP 28
#define MAXP 29
#define MINP 30
#define REDNP 31
/*************************************************************************/
/* */
/* Check whether file is open. If it is not, open it and */
/* read/write sampling information and discrete names */
/* */
/*************************************************************************/
void CheckFile(String Extension, Boolean Write)
/* --------- */
{
static char *LastExt="";
if ( ! Mf || strcmp(LastExt, Extension) )
{
LastExt = Extension;
if ( Mf )
{
fprintf(Mf, "\n");
fclose(Mf); Mf = Nil;
}
if ( Write )
{
WriteFilePrefix(Extension);
}
else
{
ReadFilePrefix(Extension);
}
}
}
/*************************************************************************/
/* */
/* Write information on system, sampling */
/* */
/*************************************************************************/
void WriteFilePrefix(String Extension)
/* --------------- */
{
Attribute Att;
time_t clock;
struct tm *now;
if ( ! (Mf = GetFile(Extension, "w")) )
{
Error(NOFILE, Fn, E_ForWrite);
}
clock = time(0);
now = localtime(&clock);
now->tm_mon++;
fprintf(Mf, "id=\"Cubist %s %d-%d%d-%d%d\"\n",
RELEASE,
now->tm_year + 1900,
now->tm_mon / 10, now->tm_mon % 10,
now->tm_mday / 10, now->tm_mday % 10);
SaveDiscreteNames();
fprintf(Mf, "prec=\"%d\" globalmean=\"%.*g\" extrap=\"%g\" insts=\"%d\" ",
Precision, PREC, GlobalMean, EXTRAP, USEINSTANCES);
if ( USEINSTANCES )
{
fprintf(Mf, "nn=\"%d\" maxd=\"%.1f\" ", NN, MAXD);
}
fprintf(Mf, "ceiling=\"%.*g\" floor=\"%.*g\"\n",
PREC, Ceiling, PREC, Floor);
/* Information for replacing missing values */
ForEach(Att, 1, MaxAtt)
{
if ( Exclude(Att) ) continue;
AsciiOut("att=", AttName[Att]);
if ( Continuous(Att) )
{
fprintf(Mf, " mean=\"%.*g\" sd=\"%.*g\" min=\"%g\" max=\"%g\"\n",
PREC, AttMean[Att], PREC, AttSD[Att],
AttMinVal[Att], AttMaxVal[Att]);
}
else
{
AsciiOut(" mode=", AttValName[Att][Max(1, Modal[Att])]);
fprintf(Mf, "\n");
}
}
/* Sampling information (if needed) */
if ( SAMPLE > 0 )
{
fprintf(Mf, "sample=\"%g\" init=\"%d\"\n", SAMPLE, KRInit);
}
if ( MEMBERS > 1 )
{
fprintf(Mf, "redn=\"%.3f\" ", ErrReduction);
}
fprintf(Mf, "entries=\"%d\"\n", MEMBERS);
}
/*************************************************************************/
/* */
/* Open model file and read header information */
/* */
/*************************************************************************/
void ReadFilePrefix(String Extension)
/* -------------- */
{
int Head;
if ( ! (Mf = GetFile(Extension, "r")) ) Error(NOFILE, Fn, "");
ReadHeader();
}
/*************************************************************************/
/* */
/* Save attribute values read with "discrete N" */
/* */
/*************************************************************************/
void SaveDiscreteNames(void)
/* ----------------- */
{
Attribute Att;
DiscrValue v;
ForEach(Att, 1, MaxAtt)
{
if ( ! StatBit(Att, DISCRETE) || MaxAttVal[Att] < 2 ) continue;
AsciiOut("att=", AttName[Att]);
AsciiOut(" elts=", AttValName[Att][2]); /* skip N/A */
ForEach(v, 3, MaxAttVal[Att])
{
AsciiOut(",", AttValName[Att][v]);
}
fprintf(Mf, "\n");
}
}
/*************************************************************************/
/* */
/* Save the current rulesets in model file */
/* */
/*************************************************************************/
void SaveCommittee(RRuleSet *Cttee, String Extension)
/* --------- */
{
int m;
CheckFile(Extension, true);
ForEach(m, 0, MEMBERS-1)
{
SaveRules(Cttee[m]);
}
}
void SaveRules(RRuleSet RS)
/* --------- */
{
int ri, d;
CRule R;
Condition C;
DiscrValue v;
Attribute Att;
Boolean First;
fprintf(Mf, "rules=\"%d\"\n", RS->SNRules);
ForEach(ri, 1, RS->SNRules)
{
R = RS->SRule[ri];
fprintf(Mf, "conds=\"%d\" cover=\"%d\" mean=\"%.*f\" "
"loval=\"%g\" hival=\"%g\" esterr=\"%.*f\"\n",
R->Size, R->Cover,
Precision+1, R->Mean,
R->LoVal, R->HiVal,
Precision+1, R->EstErr);
ForEach(d, 1, R->Size)
{
C = R->Lhs[d];
fprintf(Mf, "type=\"%d\"", C->NodeType);
AsciiOut(" att=", AttName[C->Tested]);
switch ( C->NodeType )
{
case BrDiscr:
AsciiOut(" val=", AttValName[C->Tested][C->TestValue]);
break;
case BrThresh:
if ( C->TestValue == 1 )
{
fprintf(Mf, " val=\"N/A\"");
}
else
{
fprintf(Mf, " cut=\"%.*g\" result=\"%s\"",
PREC+1, C->Cut,
( C->TestValue == 2 ? "<=" : ">" ));
}
break;
case BrSubset:
First=true;
ForEach(v, 1, MaxAttVal[C->Tested])
{
if ( In(v, C->Subset) )
{
if ( First )
{
AsciiOut(" elts=", AttValName[C->Tested][v]);
First = false;
}
else
{
AsciiOut(",", AttValName[C->Tested][v]);
}
}
}
break;
}
fprintf(Mf, "\n");
}
fprintf(Mf, "coeff=\"%.14g\"", R->Rhs[0]);
ForEach(Att, 1, MaxAtt)
{
if ( fabs(R->Rhs[Att]) > 0 )
{
AsciiOut(" att=", AttName[Att]);
fprintf(Mf, " coeff=\"%.14g\"", R->Rhs[Att]);
}
}
fprintf(Mf, "\n");
}
}
/*************************************************************************/
/* */
/* Write ASCII string with prefix, escaping any quotes */
/* */
/*************************************************************************/
void AsciiOut(String Pre, String S)
/* -------- */
{
fprintf(Mf, "%s\"", Pre);
while ( *S )
{
if ( *S == '"' || *S == '\\' ) fputc('\\', Mf);
fputc(*S++, Mf);
}
fputc('"', Mf);
}
/*************************************************************************/
/* */
/* Read the header information (id, saved names, models) */
/* */
/*************************************************************************/
void ReadHeader(void)
/* --------- */
{
Attribute Att;
DiscrValue v;
char *p, Dummy;
double Xd;
int Year, Month, Day;
/* First allocate storage for various globals */
AttMean = Alloc(MaxAtt+1, ContValue);
AttSD = Alloc(MaxAtt+1, ContValue);
AttMaxVal = Alloc(MaxAtt+1, ContValue);
AttMinVal = Alloc(MaxAtt+1, ContValue);
Modal = Alloc(MaxAtt+1, DiscrValue);
while ( true )
{
switch ( ReadProp(&Dummy) )
{
case IDP:
/* Recover year run and set base date for timestamps */
if ( sscanf(PropVal + strlen(PropVal) - 11,
"%d-%d-%d\"", &Year, &Month, &Day) == 3 )
{
SetTSBase(Year);
}
break;
case ATTP:
Unquoted = RemoveQuotes(PropVal);
Att = Which(Unquoted, AttName, 1, MaxAtt);
if ( ! Att || Exclude(Att) )
{
Error(MODELFILE, E_MFATT, Unquoted);
}
break;
case ELTSP:
/* First element ("N/A") already added by GetNames() */
MaxAttVal[Att] = 1;
for ( p = PropVal ; *p ; )
{
p = RemoveQuotes(p);
v = ++MaxAttVal[Att];
AttValName[Att][v] = Alloc(strlen(p)+1, char);
strcpy(AttValName[Att][v], p);
for ( p += strlen(p) ; *p != '"' ; p++ )
;
p++;
if ( *p == ',' ) p++;
}
AttValName[Att][MaxAttVal[Att]+1] = "";
MaxDiscrVal = Max(MaxDiscrVal, MaxAttVal[Att]+1);
break;
case PRECP:
sscanf(PropVal, "\"%d\"", &Precision);
break;
case GLOBALMEANP:
sscanf(PropVal, "\"%f\"", &GlobalMean);
break;
case EXTRAPP:
sscanf(PropVal, "\"%f\"", &EXTRAP);
break;
case INSTSP:
USEINSTANCES = PropVal[1] - '0';
if ( USEINSTANCES )
{
/* Set legacy values */
NN = 5;
MAXD = 50;
}
break;
case NNP:
sscanf(PropVal, "\"%d\"", &NN);
break;
case MAXDP:
sscanf(PropVal, "\"%f\"", &MAXD);
break;
case CEILINGP:
sscanf(PropVal, "\"%lf\"", &Xd); Ceiling = Xd;
break;
case FLOORP:
sscanf(PropVal, "\"%lf\"", &Xd); Floor = Xd;
break;
case MEANP:
sscanf(PropVal, "\"%lf\"", &Xd); AttMean[Att] = Xd;
break;
case SDP:
sscanf(PropVal, "\"%lf\"", &Xd); AttSD[Att] = Xd;
break;
case MAXP:
sscanf(PropVal, "\"%lf\"", &Xd); AttMaxVal[Att] = Xd;
break;
case MINP:
sscanf(PropVal, "\"%lf\"", &Xd); AttMinVal[Att] = Xd;
break;
case MODEP:
Unquoted = RemoveQuotes(PropVal);
Modal[Att] = Which(Unquoted,
AttValName[Att], 1, MaxAttVal[Att]);
if ( ! Modal[Att] )
{
/* An unknown modal value is an error! */
Error(MODELFILE, E_MFATTVAL, Unquoted);
}
else
if ( Modal[Att] == 1 )
{
/* This means that all training cases had value N/A.
For consistency with instance distances, this
attribute is ignored */
SpecialStatus[Att] |= SKIP;
}
break;
case SAMPLEP:
sscanf(PropVal, "\"%f\"", &SAMPLE);
break;
case INITP:
sscanf(PropVal, "\"%d\"", &KRInit);
break;
case REDNP:
sscanf(PropVal, "\"%f\"", &ErrReduction);
break;
case ENTRIESP:
sscanf(PropVal, "\"%d\"", &MEMBERS);
Entry = 0;
return;
}
}
}
/*************************************************************************/
/* */
/* Retrieve ruleset with extension Extension */
/* (Separate functions for ruleset, single rule, single condition) */
/* */
/*************************************************************************/
RRuleSet *GetCommittee(String Extension)
/* ------------- */
{
RRuleSet *Cttee;
int m;
ErrMsgs = 0;
CheckFile(Extension, false);
if ( ErrMsgs )
{
if ( Mf )
{
fclose(Mf); Mf = Nil;
}
return Nil;
}
Cttee = Alloc(MEMBERS, RRuleSet);
ForEach(m, 0, MEMBERS-1)
{
Cttee[m] = InRules();
}
fclose(Mf); Mf = Nil;
return ( ErrMsgs ? Nil : Cttee );
}
RRuleSet InRules(void)
/* ------- */
{
RRuleSet RS;
RuleNo r;
char Delim;
RS = Alloc(1, RuleSetRec);
do
{
switch ( ReadProp(&Delim) )
{
case RULESP:
sscanf(PropVal, "\"%d\"", &RS->SNRules);
break;
}
}
while ( Delim == ' ' );
/* Read each rule */
RS->SRule = Alloc(RS->SNRules+1, CRule);
ForEach(r, 1, RS->SNRules)
{
RS->SRule[r] = InRule();
RS->SRule[r]->RNo = r;
RS->SRule[r]->MNo = Entry;
}
Entry++;
return RS;
}
CRule InRule(void)
/* ------ */
{
CRule R;
int d;
char Delim;
Attribute Att=0;
float V, Range;
R = Alloc(1, RuleRec);
/* General rule information */
do
{
switch ( ReadProp(&Delim) )
{
case CONDSP:
sscanf(PropVal, "\"%d\"", &R->Size);
break;
case COVERP:
sscanf(PropVal, "\"%d\"", &R->Cover);
break;
case MEANP:
sscanf(PropVal, "\"%f\"", &R->Mean);
break;
case LOVALP:
sscanf(PropVal, "\"%f\"", &R->LoVal);
break;
case HIVALP:
sscanf(PropVal, "\"%f\"", &R->HiVal);
break;
case ESTERRP:
sscanf(PropVal, "\"%f\"", &R->EstErr);
break;
}
}
while ( Delim == ' ' );
Range = R->HiVal - R->LoVal;
R->LoLim = ( (V = R->LoVal - EXTRAP * Range) < 0 && R->LoVal >= 0 ? 0 : V );
R->HiLim = ( (V = R->HiVal + EXTRAP * Range) > 0 && R->HiVal <= 0 ? 0 : V );
/* Conditions making up rule's left-hand side */
R->Lhs = Alloc(R->Size+1, Condition);
ForEach(d, 1, R->Size)
{
R->Lhs[d] = InCondition();
}
/* Linear model associated with rule */
R->Rhs = AllocZero(MaxAtt+1, double);
do
{
switch ( ReadProp(&Delim) )
{
case ATTP:
Unquoted = RemoveQuotes(PropVal);
Att = Which(Unquoted, AttName, 1, MaxAtt);
if ( ! Att || Exclude(Att) )
{
Error(MODELFILE, E_MFATT, Unquoted);
}
break;
case COEFFP:
sscanf(PropVal, "\"%lf\"", &R->Rhs[Att]);
break;
}
}
while ( Delim == ' ' );
return R;
}
Condition InCondition(void)
/* ----------- */
{
Condition C;
char Delim;
int X;
double XD;
C = Alloc(1, CondRec);
do
{
switch ( ReadProp(&Delim) )
{
case TYPEP:
sscanf(PropVal, "\"%d\"", &X); C->NodeType = X;
break;
case ATTP:
Unquoted = RemoveQuotes(PropVal);
C->Tested = Which(Unquoted, AttName, 1, MaxAtt);
if ( ! C->Tested || Exclude(C->Tested) )
{
Error(MODELFILE, E_MFATT, Unquoted);
}
break;
case CUTP:
sscanf(PropVal, "\"%lf\"", &XD); C->Cut = XD;
break;
case RESULTP:
C->TestValue = ( PropVal[1] == '<' ? 2 : 3 );
break;
case VALP:
if ( Continuous(C->Tested) )
{
C->TestValue = 1;
}
else
{
Unquoted = RemoveQuotes(PropVal);
C->TestValue = Which(Unquoted,
AttValName[C->Tested],
1, MaxAttVal[C->Tested]);
if ( ! C->TestValue )
{
Error(MODELFILE, E_MFATTVAL, Unquoted);
}
}
break;
case ELTSP:
C->Subset = MakeSubset(C->Tested);
C->TestValue = 1;
break;
}
}
while ( Delim == ' ' );
return C;
}
/*************************************************************************/
/* */
/* ASCII reading utilities */
/* */
/*************************************************************************/
int ReadProp(char *Delim)
/* -------- */
{
int c, i;
char *p;
Boolean Quote=false;
for ( p = PropName ; (c = fgetc(Mf)) != '=' ; )
{
if ( p - PropName >= 19 || c == EOF )
{
Error(MODELFILE, E_MFEOF, "");
PropName[0] = PropVal[0] = *Delim = '\00';
return 0;
}
*p++ = c;
}
*p = '\00';
for ( p = PropVal ; ((c = fgetc(Mf)) != ' ' && c != '\n') || Quote ; )
{
if ( c == EOF )
{
Error(MODELFILE, E_MFEOF, "");
PropName[0] = PropVal[0] = '\00';
return 0;
}
if ( (i = p - PropVal) >= PropValSize )
{
Realloc(PropVal, (PropValSize += 10000) + 3, char);
p = PropVal + i;
}
*p++ = c;
if ( c == '\\' )
{
*p++ = fgetc(Mf);
}
else
if ( c == '"' )
{
Quote = ! Quote;
}
}
*p = '\00';
*Delim = c;
return Which(PropName, Prop, 1, PROPS);
}
String RemoveQuotes(String S)
/* ------------ */
{
char *p, *Start;
p = Start = S;
for ( S++ ; *S != '"' ; S++ )
{
if ( *S == '\\' ) S++;
*p++ = *S;
*S = '-';
}
*p = '\00';
return Start;
}
Set MakeSubset(Attribute Att)
/* ---------- */
{
int Bytes, b;
char *p;
Set S;
Bytes = (MaxAttVal[Att]>>3) + 1;
S = AllocZero(Bytes, unsigned char);
for ( p = PropVal ; *p ; )
{
p = RemoveQuotes(p);
b = Which(p, AttValName[Att], 1, MaxAttVal[Att]);
if ( ! b ) Error(MODELFILE, E_MFATTVAL, p);
SetBit(b, S);
for ( p += strlen(p) ; *p != '"' ; p++ )
;
p++;
if ( *p == ',' ) p++;
}
return S;
}
predict.c 0000664 0000764 0000764 00000011042 11501534415 012563 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Predict the value of a case from a ruleset */
/* ------------------------------------------ */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
/*************************************************************************/
/* */
/* Find rules that apply to case and form average */
/* */
/*************************************************************************/
float PredictValue(RRuleSet *Cttee, DataRec CaseDesc)
/* ------------ */
{
double PredSum=0;
int m;
ForEach(m, 0, MEMBERS-1)
{
PredSum += RuleSetPrediction(Cttee[m], CaseDesc);
}
return PredSum / MEMBERS;
}
float RuleSetPrediction(RRuleSet RS, DataRec CaseDesc)
/* ----------------- */
{
double Sum=0, Weight=0, Val;
int r;
CRule R;
Attribute Att;
ForEach(r, 1, RS->SNRules)
{
R = RS->SRule[r];
if ( Matches(R, CaseDesc) )
{
/* Evaluate RHS. Cannot use RawLinModel() because
have not run FindModelAtts() */
Val = R->Rhs[0];
ForEach(Att, 1, MaxAtt)
{
Val += CVal(CaseDesc, Att) * R->Rhs[Att];
}
Sum += ( Val < R->LoLim ? R->LoLim :
Val > R->HiLim ? R->HiLim : Val );
Weight += 1.0;
}
}
if ( Weight )
{
return Sum / Weight;
}
else
{
return GlobalMean;
}
}
/*************************************************************************/
/* */
/* Determine whether a case satisfies all conditions of a rule */
/* */
/*************************************************************************/
Boolean Matches(CRule R, DataRec Case)
/* ------- */
{
int d;
ForEach(d, 1, R->Size)
{
if ( ! Satisfies(Case, R->Lhs[d]) )
{
return false;
}
}
return true;
}
/*************************************************************************/
/* */
/* Evaluate a linear model on a case */
/* */
/*************************************************************************/
float RawLinModel(double *Model, DataRec Case)
/* ----------- */
{
double Sum;
Attribute Att, a;
Sum = Model[0];
ForEach(a, 1, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
Sum += Model[Att] * CVal(Case, Att);
}
return Sum;
}
float LinModel(double *Model, DataRec Case)
/* -------- */
{
float Raw;
Raw = RawLinModel(Model, Case);
return ( Raw < Floor ? Floor :
Raw > Ceiling ? Ceiling : Raw );
}
/*************************************************************************/
/* */
/* Find values predicted by a model for cases Fp to Lp. */
/* */
/*************************************************************************/
void FindPredictedValues(RRuleSet *Cttee, CaseNo Fp, CaseNo Lp)
/* ------------------- */
{
CaseNo i;
ForEach(i, Fp, Lp)
{
PredVal(Case[i]) = ( USEINSTANCES ? NNEstimate(Cttee, Case[i]) :
PredictValue(Cttee, Case[i]) );
}
}
prunetree.c 0000664 0000764 0000764 00000035760 11501534415 013157 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
double TotalErr, ExtraErr, TotalParams, ExtraParams, AdjErrLim, NewAdjErr;
Tree Weakest;
#define ResidWt PResid
/*************************************************************************/
/* */
/* Simplify tree T by successively pruning the weakest subtree. */
/* */
/* Repeat */
/* * examine each subtree of T to find the effect on global error */
/* and number of parameters if subtree pruned */
/* * if there is one with lower estimated error, prune the subtree */
/* that introduces least additional error */
/* */
/* Then, prune a little further. */
/* */
/* For each subtree S, S->Utility gives the minimum additional */
/* error obtainable by pruning S or any of its subtrees. This */
/* is used to eliminate candidates for weakest subtree. */
/* */
/*************************************************************************/
void Prune(Tree T)
/* ----- */
{
double AdjErr;
int Leaves, Stage=1;
/* Calculate numbers of parameters and coefficients */
SetProperties(T, Nil);
/* Models must be smoothed first so that error at each node
can be determined directly */
Verbosity(1, PrintTree(T, "Initial model tree"))
SmoothModels(T, Nil, 0, MaxCase);
FindErrors(T, 0, MaxCase);
Leaves = TreeLeaves(T);
/* Pruning is carried out in three stages. In the first stage, we
find a sequence of trees with non-increasing adjusted error.
The second stage allows adjusted error to increase by 0.5%,
yielding simpler models. In the final stage, pruning continues
until the number of leaves does not exceed MAXRULES */
while ( true )
{
TotalParams = T->Params;
TotalErr = T->TreeErr;
AdjErr = (MaxCase+1) * EstimateErr(TotalErr / (MaxCase+1.0),
MaxCase+1, TotalParams);
if ( Stage == 1 ) AdjErrLim = AdjErr;
Verbosity(1,
fprintf(Of, "\n(%g params, %d leaves)"
" total actual/est err %.1f/%.1f\n",
TotalParams, Leaves, TotalErr, AdjErr))
ExtraErr = 1E38; /* initialise minimum extra error */
if ( FindWeakestSubtree(T) &&
( NewAdjErr <= AdjErr || Stage > 1 ) )
{
Leaves -= TreeLeaves(Weakest) - 1;
UnsproutAndUpdate(Weakest, ExtraErr, ExtraParams);
}
else
if ( Stage == 1 )
{
Stage = 2;
AdjErrLim = AdjErr * 1.005;
}
else
{
Stage = 3;
AdjErrLim = 1E38;
}
if ( Stage == 3 && Leaves <= MAXRULES ) break;
}
}
/*************************************************************************/
/* */
/* Set Params and Coeffs for tree T. (Params currently holds */
/* the parameters equivalent of each test, but we need the */
/* totals including subtrees.) */
/* */
/* Also remove any subtrees that have very small branches */
/* */
/*************************************************************************/
void SetProperties(Tree T, Tree Parent)
/* ------------- */
{
DiscrValue v;
/* Eliminate any branches that cover too few cases */
if ( T->NodeType &&
( T->Branch[2]->Cases < MINITEMS ||
T->Branch[3]->Cases < MINITEMS ) )
{
Unsprout(T);
}
if ( T->Coeffs )
{
T->Coeffs = 0.5 * (T->Coeffs + CountCoeffs(T->Model));
}
else
{
T->Coeffs = CountCoeffs(T->Model);
}
if ( T->NodeType )
{
Parent(T) = Parent;
T->MCopy = Alloc(MaxAtt+1, double);
memcpy(T->MCopy, T->Model, (MaxAtt+1) * sizeof(double));
if ( Discrete(T->Tested) )
{
T->Params = ( T->NodeType == BrSubset ? 4 : 3 );
}
else
{
T->Params = 4;
}
ForEach(v, 1, T->Forks)
{
if ( T->Branch[v]->Cases > 0 )
{
SetProperties(T->Branch[v], T);
T->Params += T->Branch[v]->Params;
}
}
}
else
{
T->Params = T->Coeffs;
}
}
/*************************************************************************/
/* */
/* Replace a subtree by a leaf and adjust properties */
/* */
/*************************************************************************/
void UnsproutAndUpdate(Tree Pruned, double ExtraErr, double ExtraParams)
/* ----------------- */
{
DiscrValue v;
Tree T;
double MinExtraErr=1E38;
Pruned->Utility = 1E38;
for ( T = Pruned ; T ; T = Parent(T) )
{
T->TreeErr += ExtraErr;
T->Params += ExtraParams;
if ( T != Pruned )
{
ForEach(v, 1, T->Forks)
{
if ( T->Branch[v]->Utility < MinExtraErr )
{
MinExtraErr = T->Branch[v]->Utility;
}
}
T->Utility = Min(MinExtraErr, T->LeafErr - T->TreeErr);
}
}
Unsprout(Pruned);
}
/*************************************************************************/
/* */
/* Replace a subtree by a leaf */
/* */
/*************************************************************************/
void Unsprout(Tree T)
/* -------- */
{
DiscrValue v;
Progress(TreeSize(T)-1);
Verbosity(1, PrintTree(T, "Prune subtree"))
/* Free subtrees etc */
ForEach(v, 1, T->Forks)
{
FreeTree(T->Branch[v]);
}
Free(T->Branch); T->Branch = Nil;
if ( T->NodeType == BrSubset )
{
FreeVector((void **) T->Subset, 1, 3); T->Subset = Nil;
}
T->NodeType = 0;
}
/*************************************************************************/
/* */
/* Use Pregibon-like "shrinking" to smooth all models and */
/* simplify model coefficients */
/* */
/*************************************************************************/
void SmoothModels(Tree T, Tree Parent, CaseNo Fp, CaseNo Lp)
/* ------------ */
{
DiscrValue v;
Attribute Att;
CaseNo i, Bp;
CaseCount Cases;
double Wt, SumWt, SumX, SumY, SumXY, Cov,
CurrVar, ParentVar, pCurr;
/* Smooth models at subtrees */
if ( T->NodeType )
{
Bp = Fp;
ForEach(v, 1, T->Forks)
{
if ( (Cases = T->Branch[v]->Cases) > 0 )
{
SmoothModels(T->Branch[v], T, Bp, Bp + Cases - 1);
Bp += Cases;
}
}
}
/* Ignore smoothing if no variation in target or if sum of case
weights is small */
if ( CWtAtt )
{
SumWt = 0;
ForEach(i, Fp, Lp)
{
SumWt += CWeight(Case[i]);
}
}
else
{
SumWt = Lp-Fp+1;
}
/* Save errors of model in GEnv.Resid and compute variance */
CurrVar = ErrVariance(T->Model, Fp, Lp, GEnv.Resid);
if ( SumWt > 2 && CurrVar > 1E-10 )
{
for ( ; Parent ; Parent = Parent(Parent) )
{
/* Save errors of parent in GEnv.PResid and find variance */
ParentVar = ErrVariance(Parent->MCopy, Fp, Lp, GEnv.PResid);
/* Determine best combination of parent and current models
to minimise error variance. Smoothed model is
pCurr * model + (1-pCurr) * parent model
Minimisation requires covariance of error values */
SumX = SumY = SumXY = 0;
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
SumX += Wt * GEnv.Resid[i];
SumY += Wt * GEnv.PResid[i];
SumXY += Wt * GEnv.Resid[i] * GEnv.PResid[i];
}
Cov = (SumXY - SumX * SumY / SumWt) / (SumWt - 1);
pCurr = ( CurrVar + ParentVar - 2 * Cov <= 1E-12 ? 1 :
(ParentVar - Cov) / (CurrVar + ParentVar - 2 * Cov) );
/* Form smoothed model, if sensible, and recompute model
errors and error variance */
if ( pCurr > 0 && pCurr < 1 )
{
ForEach(Att, 0, MaxAtt)
{
T->Model[Att] = T->Model[Att] * pCurr +
Parent->MCopy[Att] * (1 - pCurr);
}
CurrVar = ErrVariance(T->Model, Fp, Lp, GEnv.Resid);
}
}
}
/* Simplify coefficients using AttUnit */
ForEach(Att, 1, MaxAtt)
{
if ( T->Model[Att] )
{
T->Model[Att] =
( fabs(T->Model[Att]) < MinAttCoeff(Att) ? 0 :
rint(T->Model[Att] / AttUnit[Att]) * AttUnit[Att] );
}
}
/* Adjust intercept so that weighted median error is zero and simplify */
FindModelAtts(T->Model);
ForEach(i, Fp, Lp)
{
GEnv.Resid[i] = RawLinModel(T->Model, Case[i]) - Class(Case[i]);
GEnv.ResidWt[i] = CWeight(Case[i]);
}
T->Model[0] -= MedianResid(Fp, Lp, SumWt / 2);
T->Model[0] = rint(T->Model[0] / AttUnit[0]) * AttUnit[0];
}
/*************************************************************************/
/* */
/* Save error on each case and compute variance */
/* */
/*************************************************************************/
double ErrVariance(double *Model, CaseNo Fp, CaseNo Lp, double *Err)
/* ----------- */
{
double V, SumWt=0, Sum=0, SumSq=0, Wt, Var;
CaseNo i;
FindModelAtts(Model);
ForEach(i, Fp, Lp)
{
Err[i] = V = Class(Case[i]) - LinModel(Model, Case[i]);
Wt = CWeight(Case[i]);
SumWt += Wt;
Sum += Wt * V;
SumSq += Wt * V * V;
}
return ( SumWt <= 1 ? 1E-38 :
(Var = (SumSq - Sum * Sum / SumWt) / (SumWt - 1)) > 1E-38 ? Var :
1E-38 );
}
/*************************************************************************/
/* */
/* Find current error associated with each subtree and minimum */
/* extra error if any sub-subtree is pruned */
/* */
/*************************************************************************/
void FindErrors(Tree T, CaseNo Fp, CaseNo Lp)
/* ---------- */
{
CaseNo i, Bp, Cases;
double Wt, Err=0, MinExtraErr=1E38;
DiscrValue v;
FindModelAtts(T->Model);
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
Err += Wt * fabs(Class(Case[i]) - RawLinModel(T->Model, Case[i]));
}
T->LeafErr = T->TreeErr = Err;
T->Utility = 1E38;
if ( ! T->NodeType ) return;
T->TreeErr = 0;
Bp = Fp;
ForEach(v, 1, T->Forks)
{
if ( (Cases = T->Branch[v]->Cases) > 0 )
{
FindErrors(T->Branch[v], Bp, Bp + Cases - 1);
T->TreeErr += T->Branch[v]->TreeErr;
if ( T->Branch[v]->NodeType &&
T->Branch[v]->Utility < MinExtraErr )
{
MinExtraErr = T->Branch[v]->Utility;
}
Bp += Cases;
}
}
T->Utility = Min(MinExtraErr, T->LeafErr - T->TreeErr);
}
/*************************************************************************/
/* */
/* Find "weakest" subtree in tree, defined as subtree that can */
/* be deleted with least increase in actual error while keeping */
/* adjusted error below AdjErrLim */
/* */
/*************************************************************************/
Boolean FindWeakestSubtree(Tree T)
/* ------------------ */
{
Boolean Found=false;
double ThisExtraErr, ThisAdjErr;
float ThisExtraParams;
DiscrValue v;
if ( ! T->NodeType ) return false;
/* Is this the weakest so far? */
ThisExtraErr = T->LeafErr - T->TreeErr;
ThisExtraParams = T->Coeffs - T->Params;
ThisAdjErr = (MaxCase+1) *
EstimateErr((ThisExtraErr + TotalErr) / (MaxCase+1.0),
MaxCase+1, (ThisExtraParams + TotalParams));
if ( ThisAdjErr <= AdjErrLim && ThisExtraErr < ExtraErr )
{
Weakest = T;
ExtraErr = ThisExtraErr;
ExtraParams = ThisExtraParams;
NewAdjErr = ThisAdjErr;
Found = true;
}
/* Check subtrees that could be weaker than current best */
ForEach(v, 1, T->Forks)
{
if ( T->Branch[v]->Utility <= ExtraErr )
{
Found |= FindWeakestSubtree(T->Branch[v]);
}
}
return Found;
}
/*************************************************************************/
/* */
/* Compute predicted value from model at node */
/* */
/*************************************************************************/
float TreeValue(Tree T, DataRec Case)
/* --------- */
{
DiscrValue v, f;
double Val;
Attribute Att;
Att = T->Tested;
switch ( T->NodeType )
{
case 0: /* leaf */
break;
case BrDiscr: /* test of discrete attribute */
v = DVal(Case, Att);
if ( v && v <= T->Forks && T->Branch[v]->Cases >= 1 )
{
return TreeValue(T->Branch[v], Case);
}
else
{
break;
}
case BrThresh: /* test of continuous attribute */
v = ( NotApplic(Case, Att) ? 1 :
CVal(Case, Att) <= T->Cut ? 2 : 3 );
return TreeValue(T->Branch[v], Case);
case BrSubset: /* subset test on discrete attribute */
v = DVal(Case, Att);
f = ( v == 1 ? 1 : In(v, T->Subset[2]) ? 2 : 3 );
return TreeValue(T->Branch[f], Case);
}
/* Calculate value at this node. Cannot use LinModel() */
Val = T->Model[0];
ForEach(Att, 1, MaxAtt)
{
Val += CVal(Case, Att) * T->Model[Att];
}
return ( Val < Floor ? Floor : Val > Ceiling ? Ceiling : Val );
}
/*************************************************************************/
/* */
/* Find median of residuals of cases in GEnv.Filtered. */
/* Want initially contains half of the weight of cases Fp..Lp */
/* */
/*************************************************************************/
#define MSwap(a,b) {Hold = GEnv.Resid[a];\
GEnv.Resid[a] = GEnv.Resid[b];\
GEnv.Resid[b] = Hold;\
Hold = GEnv.ResidWt[a];\
GEnv.ResidWt[a] = GEnv.ResidWt[b];\
GEnv.ResidWt[b] = Hold;}
float MedianResid(CaseNo Fp, CaseNo Lp, double Want)
/* ----------- */
{
CaseNo i, Middle, High;
double LowWt, MiddleWt, Hold, Thresh, Val;
while ( Fp < Lp )
{
Thresh = GEnv.Resid[(Fp+Lp) / 2];
/* Divide cases into three groups:
Fp .. Middle-1: values < Thresh
Middle .. High: values = Thresh
High+1 .. Lp: values > Thresh */
LowWt = MiddleWt = 0;
for ( Middle = Fp ; GEnv.Resid[Middle] < Thresh ; Middle++ )
{
LowWt += GEnv.ResidWt[Middle];
}
for ( High = Lp ; GEnv.Resid[High] > Thresh ; High-- )
;
for ( i = Middle ; i <= High ; )
{
if ( (Val = GEnv.Resid[i]) < Thresh )
{
LowWt += GEnv.ResidWt[i];
MSwap(i, Middle);
Middle++;
i++;
}
else
if ( Val > Thresh )
{
MSwap(i, High);
High--;
}
else
{
MiddleWt += GEnv.ResidWt[i];
i++;
}
}
if ( Want <= LowWt )
{
Lp = Middle-1;
}
else
if ( Want > LowWt + MiddleWt )
{
Want -= LowWt + MiddleWt;
Fp = High+1;
}
else
{
return Thresh;
}
}
return GEnv.Resid[Fp];
}
regress.c 0000664 0000764 0000764 00000034442 11501534415 012614 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Find best-fit linear model by matrix inversion */
/* ---------------------------------------------- */
/* This code includes extensions to the standard methodology, */
/* principally model simplification and outlier elimination */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
/* x[i][v] = data (in Case) */
/* y[i] = dependent var (in Class) */
/* least-squares solution given by xTx.Model = xTy */
/*************************************************************************/
/* */
/* Find regression model for items Fp through Lp. */
/* Do not use variables in GEnv.DoNotUse. */
/* */
/*************************************************************************/
void Regress(CaseNo Fp, CaseNo Lp, double *Model)
/* ------- */
{
CaseNo i, Kp;
int j, jj, k, kk;
ContValue ClassVal, JVal;
double Val, SumR=0, Bias=0, AvResid, Wt, Cases=0;
Attribute a, Att;
Boolean First;
/* Find candidate attributes */
GEnv.NModelAtt = 0;
GEnv.ModelAtt[0] = 0;
GEnv.Mean[0] = GEnv.Var[0] = Model[0] = GEnv.AvDev[0] = 0;
GEnv.ZeroCoeff[0] = false;
ForEach(Att, 1, MaxAtt)
{
if ( Continuous(Att) && Att != ClassAtt && ! Skip(Att) &&
( ! GEnv.DoNotUse || ! GEnv.DoNotUse[Att] ) )
{
GEnv.ModelAtt[++GEnv.NModelAtt] = Att;
GEnv.Mean[Att] = GEnv.Var[Att] = Model[Att] = GEnv.AvDev[Att] = 0;
GEnv.ZeroCoeff[Att] = false;
}
else
{
GEnv.ZeroCoeff[Att] = true;
}
}
/* Find means and variances in one data pass */
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
Cases += Wt;
ForEach(a, 0, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
if ( a > 0 && NotApplic(Case[i], Att) )
{
GEnv.ZeroCoeff[Att] = true;
GEnv.ModelAtt[a--] = GEnv.ModelAtt[GEnv.NModelAtt--];
}
else
{
GEnv.Mean[Att] += Wt * (Val = CVal(Case[i], Att));
GEnv.Var[Att] += Wt * Val * Val;
}
}
}
GEnv.Mean[0] /= Cases;
ForEach(a, 1, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
GEnv.Mean[Att] /= Cases;
GEnv.Var[Att] = (GEnv.Var[Att] - Cases * GEnv.Mean[Att]
* GEnv.Mean[Att]) / (Cases-1);
if ( GEnv.Var[Att] < 1E-6 )
{
GEnv.ZeroCoeff[Att] = true;
GEnv.ModelAtt[a--] = GEnv.ModelAtt[GEnv.NModelAtt--];
}
}
/* Now find average deviations in another data pass */
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
ForEach(a, 0, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
GEnv.AvDev[Att] += Wt * fabs(CVal(Case[i], Att) - GEnv.Mean[Att]);
}
}
ForEach(a, 0, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
GEnv.AvDev[Att] /= Cases;
}
/* Don't attempt to form model if there are no possible coefficients
or if items < 2 * possible coefficients */
if ( GEnv.NModelAtt < 1 || Cases < 2 * GEnv.NModelAtt )
{
Model[0] = GEnv.Mean[0];
ForEach(Att, 1, MaxAtt)
{
Model[Att] = 0;
}
return;
}
/* Form linear model by matrix inversion */
BuildTables(Fp, Lp);
Solve(Model);
SimplifyModel(Case, Fp, Lp, Model);
/* See whether ought to exclude outliers and try again.
An outlier wrt the linear model is a case whose residual
error magnitude is more than five times the average */
FindModelAtts(Model);
if ( ! GEnv.NModelAtt ) return; /* just a mean */
/* Compute the residuals and correct any bias caused by limiting
predictions to range [Floor, Ceiling] */
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
GEnv.Resid[i] = Class(Case[i]) - LinModel(Model, Case[i]);
SumR += Wt * fabs(GEnv.Resid[i]);
Bias += Wt * GEnv.Resid[i];
}
AvResid = SumR / Cases;
Model[0] += Bias / Cases;
Kp = Fp-1;
First = true;
ForEach(i, Fp, Lp)
{
if ( fabs(GEnv.Resid[i]) > 5 * AvResid )
{
/* If this is the first outlier, re-initialise active
attributes and preserve case order */
if ( First )
{
FindActiveAtts();
First = false;
}
/* Remove contribution of this case */
Wt = CWeight(Case[i]);
GEnv.xTx[0][0] -= Wt;
GEnv.xTy[0] -= Wt * (ClassVal = Class(Case[i]));
ForEach(jj, 1, GEnv.NModelAtt)
{
j = GEnv.ModelAtt[jj];
GEnv.xTy[j] -= Wt * (JVal = CVal(Case[i], j)) * ClassVal;
GEnv.xTx[j][0] -= Wt * JVal;
ForEach(kk, 1, jj)
{
k = GEnv.ModelAtt[kk];
GEnv.xTx[j][k] -= Wt * JVal * CVal(Case[i], k);
}
}
}
else
{
/* Keep this case */
GEnv.Filtered[++Kp] = Case[i];
}
}
/* Rebuild if there are outliers. */
if ( Kp < Lp )
{
Verbosity(1, fprintf(Of, "(Exclude %d outliers)\n", Lp - Kp))
/* Construct new model with outliers excluded. Tables
GEnv.xTx and GEnv.xTy do not need to be regenerated */
Solve(Model);
SimplifyModel(GEnv.Filtered, Fp, Kp, Model);
}
}
/*************************************************************************/
/* */
/* Scan data to form GEnv.xTx and GEnv.xTy */
/* */
/*************************************************************************/
void BuildTables(CaseNo Fp, CaseNo Lp)
/* ----------- */
{
int i, j, jj, k, kk;
ContValue ClassVal, JVal;
double Wt;
FindActiveAtts();
/* Initialise GEnv.xTx and GEnv.xTy */
ForEach(jj, 0, GEnv.NModelAtt)
{
j = GEnv.ModelAtt[jj];
GEnv.xTy[j] = 0;
ForEach(kk, 0, jj)
{
GEnv.xTx[j][GEnv.ModelAtt[kk]] = 0;
}
}
/* Evaluate GEnv.xTx[j][k] etc. for k <= j; j=0 and k=0 are
handled as special cases.
(This loop has been reorganised to minimize cache misses
that seem to be very important for Suns and AMD Athlons.) */
GEnv.xTx[0][0] = 0;
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
GEnv.xTx[0][0] += Wt;
GEnv.xTy[0] += Wt * (ClassVal = Class(Case[i]));
ForEach(jj, 1, GEnv.NModelAtt)
{
j = GEnv.ModelAtt[jj];
GEnv.xTy[j] += Wt * (JVal = CVal(Case[i], j)) * ClassVal;
GEnv.xTx[j][0] += Wt * JVal;
ForEach(kk, 1, jj)
{
k = GEnv.ModelAtt[kk];
GEnv.xTx[j][k] += Wt * JVal * CVal(Case[i], k);
}
}
}
}
/*************************************************************************/
/* */
/* Solve normal equations by matrix inversion. */
/* Tables xTx and xTy are copied since we will need them again */
/* during model simplification and if outliers are excluded */
/* */
/*************************************************************************/
void Solve(double *Model)
/* ----- */
{
int j, k, jj, max;
double Pivot, MinPivot, MaxElt, Try;
Boolean Singular=false;
if ( ! GEnv.NModelAtt )
{
Model[0] = GEnv.xTy[0] / GEnv.xTx[0][0];
ForEach(j, 1, MaxAtt)
{
Model[j] = 0;
}
return;
}
/* Set up (destructible) copies A and B using only active attributes */
ForEach(j, 0, GEnv.NModelAtt)
{
ForEach(k, 0, j)
{
GEnv.A[j][k] = GEnv.A[k][j] =
GEnv.xTx[GEnv.ModelAtt[j]][GEnv.ModelAtt[k]];
}
GEnv.B[j] = GEnv.xTy[GEnv.ModelAtt[j]];
}
/* Construct inverse */
ForEach(j, 0, GEnv.NModelAtt)
{
/* Find best pivot for column j. First, find max value in
original column j, and use that to estimate minimum
acceptable pivot = maximum expected rounding error */
jj = GEnv.ModelAtt[j];
MaxElt = 0;
ForEach(k, 0, j)
{
if ( (Try = fabs(GEnv.xTx[GEnv.ModelAtt[k]][jj])) > MaxElt )
{
MaxElt = Try;
}
}
MinPivot = (GEnv.NModelAtt+1) * MaxElt * 1E-12;
max = j;
Pivot = fabs(GEnv.A[j][j]);
ForEach(k, j+1, GEnv.NModelAtt)
{
if ( (Try = fabs(GEnv.A[k][j])) > Pivot )
{
max = k;
Pivot = Try;
}
}
/* Make diagonal element 1 if possible. If cannot, note that
matrix is singular and exclude this attribute */
if ( Pivot < MinPivot )
{
GEnv.ZeroCoeff[GEnv.ModelAtt[j]] = Singular = true;
Verbosity(3,
printf("=== no pivot for %s (%g, max elt=%g)\n",
AttName[GEnv.ModelAtt[j]], Pivot, MaxElt))
}
else
{
if ( max != j )
{
ExchangeRow(GEnv.B, max, j);
}
Pivot = GEnv.A[j][j]; /* actual pivot might be -ve */
ForEach(k, j, GEnv.NModelAtt)
{
GEnv.A[j][k] /= Pivot;
}
GEnv.B[j] /= Pivot;
ForEach(k, j+1, GEnv.NModelAtt)
{
AddRow(GEnv.B, j, k, -GEnv.A[k][j]);
}
}
}
/* If matrix is singular, try again without excluded attributes */
if ( Singular )
{
FindActiveAtts();
Solve(Model);
return;
}
/* Matrix is now triangular with 1 along diagonal */
for ( j = GEnv.NModelAtt-1 ; j >= 0 ; j-- )
{
ForEach(k, j+1, GEnv.NModelAtt)
{
GEnv.B[j] -= GEnv.A[j][k] * GEnv.B[k];
}
}
/* Transfer values into Model */
ForEach(j, 1, MaxAtt)
{
Model[j] = 0;
}
ForEach(j, 0, GEnv.NModelAtt)
{
Model[GEnv.ModelAtt[j]] = GEnv.B[j];
}
}
/*************************************************************************/
/* */
/* Construct list of active (non-excluded) attributes */
/* */
/*************************************************************************/
void FindActiveAtts(void)
/* -------------- */
{
Attribute Att;
GEnv.NModelAtt = 0;
ForEach(Att, 0, MaxAtt)
{
if ( ! GEnv.ZeroCoeff[Att] ) GEnv.ModelAtt[GEnv.NModelAtt++] = Att;
}
GEnv.NModelAtt--;
}
/*************************************************************************/
/* */
/* Add Factor * row From to row To */
/* */
/*************************************************************************/
void AddRow(double *Model, short From, short To, double Factor)
/* ------ */
{
short col;
double *ATo, *AFrom;
ATo = &GEnv.A[To][0];
AFrom = &GEnv.A[From][0];
ForEach(col, 0, GEnv.NModelAtt)
{
ATo[col] += Factor * AFrom[col];
}
Model[To] += Factor * Model[From];
}
/*************************************************************************/
/* */
/* Exchange rows From and To */
/* */
/*************************************************************************/
void ExchangeRow(double *Model, short From, short To)
/* ----------- */
{
short col;
double HoldVal;
ForEach(col, 0, GEnv.NModelAtt)
{
HoldVal = GEnv.A[From][col];
GEnv.A[From][col] = GEnv.A[To][col];
GEnv.A[To][col] = HoldVal;
}
HoldVal = Model[From];
Model[From] = Model[To];
Model[To] = HoldVal;
}
/*************************************************************************/
/* */
/* Count coefficients in Model */
/* */
/*************************************************************************/
int CountCoeffs(double *Model)
/* ----------- */
{
int Number=1;
Attribute Att;
if ( Model )
{
ForEach(Att, 1, MaxAtt)
{
if ( fabs(Model[Att]) > 0 )
{
Number++;
}
}
}
return Number;
}
/*************************************************************************/
/* */
/* Simplify a model. */
/* Drop coefficients one at a time, computing adjusted error; */
/* then pick the model with lowest adjusted error. */
/* */
/*************************************************************************/
void SimplifyModel(DataRec *D, CaseNo Fp, CaseNo Lp, double *Model)
/* ------------- */
{
double ModErr, AdjModErr, BestAdjErr=1E10, Contrib, LeastContrib,
Wt, Cases=0;
CaseNo i;
Attribute Att, Drop;
int a;
Boolean Stable;
ForEach(i, Fp, Lp)
{
Wt = CWeight(D[i]);
Cases += Wt;
}
/* Drop coefficients one at a time, selecting the coefficient that
makes the least contribution to the model */
memcpy(GEnv.SaveZero, GEnv.ZeroCoeff, (MaxAtt+1)*sizeof(Boolean));
do
{
Verbosity(3,
{
fprintf(Of, "Model %.3f", Model[0]);
ForEach(Att, 1, MaxAtt)
{
if ( GEnv.ZeroCoeff[Att] )
{
fprintf(Of, " *");
}
else
{
fprintf(Of, " %.3f", Model[Att]);
}
}
fprintf(Of, "\n");
})
Stable = true;
Drop = 0;
ForEach(a, 1, GEnv.NModelAtt)
{
Att = GEnv.ModelAtt[a];
Contrib = fabs(Model[Att] * GEnv.AvDev[Att]);
if ( ! Drop || Contrib < LeastContrib )
{
Drop = Att;
LeastContrib = Contrib;
}
if ( Contrib > 1000 * GEnv.AvDev[0] ) Stable = false;
}
/* Check whether current model is best so far */
if ( Stable && Cases >= 2 * GEnv.NModelAtt )
{
ModErr = AverageErr(D, Fp, Lp, Model);
AdjModErr = EstimateErr(ModErr, Cases, GEnv.NModelAtt);
Verbosity(3,
fprintf(Of, "Cases %d:%d mod err=%.2f/%.2f%s\n",
Fp, Lp, ModErr, AdjModErr,
( AdjModErr <= BestAdjErr ? " **" : "" )))
if ( AdjModErr <= BestAdjErr )
{
BestAdjErr = AdjModErr;
memcpy(GEnv.BestModel, Model, (MaxAtt+1)*sizeof(double));
}
}
if ( Drop )
{
Verbosity(3,
fprintf(Of, "Drop %s coeff=%.3f mean=%.3f |dev|=%.3f\n",
AttName[Drop], Model[Drop], GEnv.Mean[Drop],
GEnv.AvDev[Drop]))
Model[Drop] = 0;
GEnv.ZeroCoeff[Drop] = true;
/* Construct new model using remaining attributes */
FindActiveAtts();
Solve(Model);
}
}
while ( Drop );
memcpy(Model, GEnv.BestModel, (MaxAtt+1)*sizeof(double));
memcpy(GEnv.ZeroCoeff, GEnv.SaveZero, (MaxAtt+1)*sizeof(Boolean));
}
rules.c 0000664 0000764 0000764 00000027156 11501534415 012300 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Miscellaneous routines for rule handling */
/* ---------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
/*************************************************************************/
/* */
/* Add a new rule to the current ruleset, by updating Rule[], */
/* NRules and, if necessary, RuleSpace */
/* */
/*************************************************************************/
Boolean NewRule(Condition Cond[], int NCond, Boolean *Deleted,
CaseCount Cover, float Mean, float LoVal, float HiVal,
float EstErr, double *Model)
/* ------- */
{
int d, dd, id, r, Size=0, Bytes;
CRule R;
Condition *Lhs;
Boolean Exclude=false;
float Range, V;
extern double *Total;
/* Sort and copy the conditions */
ForEach(d, 1, NCond)
{
if ( ! Deleted[d] ) Size++;
}
Lhs = Alloc(Size+1, Condition);
ForEach(d, 1, Size)
{
dd = 0;
ForEach(id, 1, NCond)
{
if ( ! Deleted[id] &&
( ! dd || Total[id] > Total[dd] ) )
{
dd = id;
}
}
Lhs[d] = Alloc(1, CondRec);
memcpy(Lhs[d], Cond[dd], sizeof(CondRec));
if ( Lhs[d]->NodeType == BrSubset )
{
Bytes = (MaxAttVal[Lhs[d]->Tested]>>3) + 1;
Lhs[d]->Subset = Alloc(Bytes, unsigned char);
memcpy(Lhs[d]->Subset, Cond[dd]->Subset, Bytes);
}
Deleted[dd] = true;
}
/* See if rule already exists */
for ( r = 1 ; ! Exclude && r <= NRules ; r++ )
{
if ( SameRule(r, Lhs, Size) )
{
Verbosity(1, fprintf(Of, "\tduplicates rule %d\n", r))
Exclude = true;
/* Save this model if has lower estimated error */
if ( EstErr < Rule[r]->EstErr )
{
memcpy(Rule[r]->Rhs, Model, (MaxAtt+1)*sizeof(double));
Rule[r]->EstErr = EstErr;
}
}
}
if ( Exclude )
{
/* Free any subsets */
ForEach(d, 1, Size)
{
if ( Lhs[d]->NodeType == BrSubset )
{
FreeUnlessNil(Lhs[d]->Subset);
}
}
FreeVector((void **) Lhs, 1, Size);
return false;
}
/* Make sure there is enough room for the new rule */
NRules++;
if ( NRules >= RuleSpace )
{
RuleSpace += 100;
if ( RuleSpace > 100 )
{
Realloc(Rule, RuleSpace, CRule);
}
else
{
Rule = Alloc(RuleSpace, CRule);
}
}
/* Form the new rule */
Rule[NRules] = R = Alloc(1, RuleRec);
R->RNo = NRules;
R->Size = Size;
R->Lhs = Lhs;
R->Cover = Cover;
R->Mean = Mean;
R->LoVal = LoVal;
R->HiVal = HiVal;
Range = HiVal - LoVal;
R->LoLim = ( (V = R->LoVal - EXTRAP * Range) < 0 && R->LoVal >= 0 ? 0 : V );
R->HiLim = ( (V = R->HiVal + EXTRAP * Range) > 0 && R->HiVal <= 0 ? 0 : V );
R->Rhs = AllocZero(MaxAtt+1, double);
memcpy(R->Rhs, Model, (MaxAtt+1)*sizeof(double));
R->EstErr = EstErr;
Verbosity(1, PrintRule(R))
return true;
}
/*************************************************************************/
/* */
/* Decide whether the given rule duplicates rule r */
/* */
/*************************************************************************/
Boolean SameRule(RuleNo r, Condition Cond[], int NConds)
/* -------- */
{
int d, i, Bytes;
if ( Rule[r]->Size != NConds )
{
return false;
}
ForEach(d, 1, NConds)
{
if ( Rule[r]->Lhs[d]->NodeType != Cond[d]->NodeType ||
Rule[r]->Lhs[d]->Tested != Cond[d]->Tested )
{
return false;
}
switch ( Cond[d]->NodeType )
{
case BrDiscr:
if ( Rule[r]->Lhs[d]->TestValue != Cond[d]->TestValue )
{
return false;
}
break;
case BrThresh:
if ( Rule[r]->Lhs[d]->TestValue != Cond[d]->TestValue ||
Rule[r]->Lhs[d]->Cut != Cond[d]->Cut )
{
return false;
}
break;
case BrSubset:
Bytes = (MaxAttVal[Cond[d]->Tested]>>3) + 1;
ForEach(i, 0, Bytes-1)
{
if ( Rule[r]->Lhs[d]->Subset[i] != Cond[d]->Subset[i] )
{
return false;
}
}
}
}
return true;
}
/*************************************************************************/
/* */
/* Free space occupied by a rule or a committee */
/* */
/*************************************************************************/
void ReleaseRule(CRule R)
/* ----------- */
{
int d;
ForEach(d, 1, R->Size)
{
if ( R->Lhs[d]->NodeType == BrSubset )
{
FreeUnlessNil(R->Lhs[d]->Subset);
}
FreeUnlessNil(R->Lhs[d]);
}
FreeUnlessNil(R->Lhs);
FreeUnlessNil(R->Rhs);
FreeUnlessNil(R);
}
void FreeCttee(RRuleSet *Cttee)
/* --------- */
{
int m, r;
RRuleSet RS;
ForEach(m, 0, MEMBERS-1)
{
if ( ! (RS = Cttee[m]) ) continue;
ForEach(r, 1, RS->SNRules)
{
ReleaseRule(RS->SRule[r]);
}
Free(RS->SRule);
Free(RS);
}
Free(Cttee);
}
/*************************************************************************/
/* */
/* Print a ruleset */
/* */
/*************************************************************************/
void PrintRules(RRuleSet RS, String Msg)
/* ---------- */
{
int r;
fprintf(Of, "\n%s\n", Msg);
ForEach(r, 1, RS->SNRules)
{
PrintRule(RS->SRule[r]);
}
}
/*************************************************************************/
/* */
/* Print the rule R */
/* */
/*************************************************************************/
void PrintRule(CRule R)
/* --------- */
{
int c, d, dd, id, LineLen, EntryLen, Indent, NCoeff=0;
Attribute Att;
char Entry[1000];
double *Model;
float *Importance;
if ( MEMBERS > 1 )
{
fprintf(Of, "\n " T_Rule " %d/%d", R->MNo+1, R->RNo);
}
else
{
fprintf(Of, "\n " T_Rule " %d", R->RNo);
}
fprintf(Of, TX_RInfo(R->Cover, Precision+1, R->Mean,
R->LoVal, R->HiVal, R->EstErr));
if ( R->Size )
{
fprintf(Of, " " T_If "\n");
/* Mark all conditions as unprinted or-ing flag to NodeType */
ForEach(d, 1, R->Size)
{
R->Lhs[d]->NodeType |= 8;
}
ForEach(d, 1, R->Size)
{
dd = 0;
ForEach(id, 1, R->Size)
{
if ( (R->Lhs[id]->NodeType & 8) &&
( ! dd || Before(R->Lhs[id], R->Lhs[dd]) ) )
{
dd = id;
}
}
R->Lhs[dd]->NodeType &= 7;
PrintCondition(R->Lhs[dd]);
}
fprintf(Of, " " T_Then "\n");
}
/* Print the model. First estimate the importance of the coefficients */
Model = R->Rhs;
Importance = AllocZero(MaxAtt+1, float);
ForEach(Att, 1, MaxAtt)
{
if ( Att != ClassAtt && Model[Att] )
{
Importance[Att] = fabs(Model[Att]) * AttSD[Att];
NCoeff++;
}
}
sprintf(Entry, "%s =", AttName[ClassAtt]);
Indent = CharWidth(Entry);
sprintf(Entry + Indent, " %.14g", Model[0]);
fprintf(Of, "\t%s", Entry);
LineLen = CharWidth(Entry);
ForEach(c, 1, NCoeff)
{
/* Select the next attribute to print */
Att = 1;
ForEach(d, 2, MaxAtt)
{
if ( Importance[d] > Importance[Att] ) Att = d;
}
Importance[Att] = 0;
/* Print, breaking lines when necessary */
sprintf(Entry, " %c %.14g %s",
( Model[Att] > 0 ? '+' : '-' ),
fabs(Model[Att]),
AttName[Att]);
EntryLen = CharWidth(Entry);
if ( LineLen + EntryLen > 72 )
{
fprintf(Of, "\n\t%*s", Indent, " ");
LineLen = Indent;
}
fprintf(Of, "%s", Entry);
LineLen += EntryLen;
}
fprintf(Of, "\n");
Free(Importance);
}
/*************************************************************************/
/* */
/* Print a condition C of a rule */
/* */
/*************************************************************************/
void PrintCondition(Condition C)
/* -------------- */
{
DiscrValue v, pv, Last, Values=0;
Boolean First=true;
Attribute Att;
int Col, Base, Entry;
char CVS[20];
v = C->TestValue;
Att = C->Tested;
fprintf(Of, "\t%s", AttName[Att]);
if ( v < 0 )
{
fprintf(Of, T_IsUnknown);
return;
}
switch ( C->NodeType )
{
case BrDiscr:
fprintf(Of, " = %s\n", AttValName[Att][v]);
break;
case BrThresh:
if ( v == 1 )
{
fprintf(Of, " = N/A\n");
}
else
{
CValToStr(C->Cut, Att, CVS);
fprintf(Of, " %s %s\n", ( v == 2 ? "<=" : ">" ), CVS);
}
break;
case BrSubset:
/* Count values at this branch */
ForEach(pv, 1, MaxAttVal[Att])
{
if ( In(pv, C->Subset) )
{
Last = pv;
Values++;
}
}
if ( Values == 1 )
{
fprintf(Of, " = %s\n", AttValName[Att][Last]);
break;
}
if ( Ordered(Att) )
{
/* Find first value */
for ( pv = 1 ; ! In(pv, C->Subset) ; pv++ )
;
fprintf(Of, " " T_InRange " [%s-%s]\n",
AttValName[Att][pv], AttValName[Att][Last]);
break;
}
/* Must keep track of position to break long lines */
fprintf(Of, " %s {", T_ElementOf);
Col = Base = CharWidth(AttName[Att]) + CharWidth(T_ElementOf) + 11;
ForEach(pv, 1, MaxAttVal[Att])
{
if ( In(pv, C->Subset) )
{
Entry = CharWidth(AttValName[Att][pv]);
if ( First )
{
First = false;
}
else
if ( Col + Entry + 2 >= Width )
{
Col = Base;
fprintf(Of, ",\n%*s", Col, "");
}
else
{
fprintf(Of, ", ");
Col += 2;
}
fprintf(Of, "%s", AttValName[Att][pv]);
Col += Entry;
}
}
fprintf(Of, "}\n");
}
}
/*************************************************************************/
/* */
/* Check whether a case satisfies a condition */
/* */
/*************************************************************************/
Boolean Satisfies(DataRec CaseDesc, Condition OneCond)
/* --------- */
{
DiscrValue v;
ContValue cv;
DiscrValue Outcome;
Attribute Att;
Att = OneCond->Tested;
/* Determine the outcome of this test on this item */
switch ( OneCond->NodeType )
{
case BrDiscr: /* test of discrete attribute */
v = DVal(CaseDesc, Att);
Outcome = ( v == 0 ? -1 : v );
break;
case BrThresh: /* test of continuous attribute */
cv = CVal(CaseDesc, Att);
Outcome = ( NotApplic(CaseDesc, Att) ? 1 :
cv <= OneCond->Cut ? 2 : 3 );
break;
case BrSubset: /* subset test on discrete attribute */
v = DVal(CaseDesc, Att);
Outcome = ( v <= MaxAttVal[Att] && In(v, OneCond->Subset) ?
OneCond->TestValue : 0 );
}
return ( Outcome == OneCond->TestValue );
}
sort.c 0000664 0000764 0000764 00000005016 11501534415 012124 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Sorting utilities */
/* ----------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define SwapSRec(a,b) {Xab=SRec[a]; SRec[a]=SRec[b]; SRec[b]=Xab;}
/*************************************************************************/
/* */
/* Sort elements Fp to Lp of SRec */
/* */
/*************************************************************************/
void Cachesort(CaseNo Fp, CaseNo Lp)
/* --------- */
{
CaseNo i, Middle, High;
ContValue Thresh, Val;
SortRec Xab;
while ( Fp < Lp )
{
Thresh = SRec[(Fp+Lp) / 2].V;
/* Divide elements into three groups:
Fp .. Middle-1: values < Thresh
Middle .. High: values = Thresh
High+1 .. Lp: values > Thresh */
for ( Middle = Fp ; SRec[Middle].V < Thresh ; Middle++ )
;
for ( High = Lp ; SRec[High].V > Thresh ; High-- )
;
for ( i = Middle ; i <= High ; )
{
if ( (Val = SRec[i].V) < Thresh )
{
SwapSRec(Middle, i);
Middle++;
i++;
}
else
if ( Val > Thresh )
{
SwapSRec(High, i);
High--;
}
else
{
i++;
}
}
/* Sort the first group */
Cachesort(Fp, Middle-1);
/* Continue with the last group */
Fp = High+1;
}
}
stats.c 0000664 0000764 0000764 00000004601 11501534415 012272 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
double AverageDev(float Mean, CaseNo Fp, CaseNo Lp)
/* ---------- */
{
double Wt, SumWt=0, Sum=0;
CaseNo i;
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
SumWt += Wt;
Sum += Wt * fabs(Mean - Class(Case[i]));
}
return Sum / SumWt;
}
double SD(double N, double Sum, double SumSq)
/* -- */
{
return ( N < 2 ? GlobalSD :
sqrt( (SumSq - Sum * Sum / N + 1E-3) / (N - 1) ) );
}
double AverageErr(DataRec *D, CaseNo Fp, CaseNo Lp, double *Model)
/* ---------- */
{
CaseNo i;
double Wt, Sum=0, SumWt=0;
FindModelAtts(Model);
ForEach(i, Fp, Lp)
{
Wt = CWeight(Case[i]);
Sum += Wt * fabs(Class(D[i]) - LinModel(Model, D[i]));
SumWt += Wt;
}
return Sum / SumWt;
}
double ComputeGain(Tree Node)
/* ----------- */
{
double Resid=0, Cases=0;
DiscrValue v;
ForEach(v, 1, 3)
{
Cases += GEnv.BrFreq[v];
Resid += GEnv.BrFreq[v] *
SD(GEnv.BrFreq[v], GEnv.BrSum[v], GEnv.BrSumSq[v]);
}
return Node->SD - Resid / Cases;
}
double EstimateErr(double Val, double NData, float NParam)
/* ----------- */
{
if ( NParam >= NData ) NParam = NData - 1;
return Val * (NData + NParam) / (NData - NParam);
}
summary.c 0000664 0000764 0000764 00000012046 11501534415 012633 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Show summary of performance on test data */
/* ---------------------------------------- */
/* */
/*************************************************************************/
#include
#include
#include
#define ForEach(v,f,l) for(v=f;v<=l;v++)
#define qsqrt(x) ((x)>0? sqrt((float) x): 0.0)
#define Plot(v) (Scale*(v-Base)+2)
#define Max(a,b) (a > b ? a : b)
/*************************************************************************/
/* */
/* This is a specialised form of the getopt utility. */
/* */
/*************************************************************************/
char *OptArg, *Option;
char ProcessOption(int Argc, char *Argv[], char *Options)
/* ------------- */
{
int i;
static int OptNo=1;
if ( OptNo >= Argc ) return '\00';
if ( *(Option = Argv[OptNo++]) != '-' ) return '?';
for ( i = 0 ; Options[i] ; i++ )
{
if ( Options[i] == Option[1] )
{
OptArg = ( Options[i+1] != '+' ? 0 :
Option[2] ? Option+2 :
OptNo < Argc ? Argv[OptNo++] : "0" );
return Option[1];
}
}
return '?';
}
int main(int Argc, char *Argv[])
/* ---- */
{
int o;
char *FileStem=0, FileName[100], Msg[100], Show=1, Quit=0;
FILE *F;
int i, N=0, Size=0, Precision=0, Denom, Digits;
float *RVal, *PVal, MinRVal, MaxRVal, MinPVal, MaxPVal, Base, Scale,
V, Mean, Real, Pred;
double SumX=0, SumY=0, SumXY=0, SumXX=0, SumYY=0,
SumAbsErr=0, SumBaseAbsErr=0;
char Line[200], AvAbsErr[20], CorCoeff[20], RelErr[20];
/* Process options */
while ( (o = ProcessOption(Argc, Argv, "f+n")) )
{
switch (o)
{
case 'f': sprintf(FileName, "%s.pred", OptArg);
FileStem = OptArg;
break;
case 'n': Show = 0;
break;
case '?': ; /* ignore illegal options */
}
}
/* Open file containing output */
if ( ! FileStem )
{
sprintf(Msg, "Result file not specified");
Quit = 1;
}
else
if ( ! (F = fopen(FileName, "r")) )
{
sprintf(Msg, "Cannot open file %s", FileName);
Quit = 1;
}
if ( Quit )
{
{
fprintf(stderr, "%s\n", Msg);
}
exit(1);
}
while ( fgets(Line, 200, F) )
{
if ( sscanf(Line, "(Default value %g", &Mean) == 1 )
;
else
if ( sscanf(Line, "%f %f", &Real, &Pred) == 2 )
{
if ( N >= Size )
{
Size += 1000;
if ( Size > 1000 )
{
RVal = (float *) realloc(RVal, Size * sizeof(float));
PVal = (float *) realloc(PVal, Size * sizeof(float));
}
else
{
RVal = (float *) malloc(Size * sizeof(float));
PVal = (float *) malloc(Size * sizeof(float));
}
}
RVal[N] = Real;
PVal[N] = Pred;
N++;
SumAbsErr += fabs(Real - Pred);
SumBaseAbsErr += fabs(Real - Mean);
SumX += Real;
SumXX += Real * Real;
SumY += Pred;
SumYY += Pred * Pred;
SumXY += Real * Pred;
/* Determine max precision so far */
V = fabs(Real - rint(Real));
Denom = 10000;
Digits = 4;
while ( Digits >= 0 &&
fabs(((int) (V * Denom + 0.5)) - V * Denom) < 0.1 )
{
Denom /= 10;
Digits--;
}
if ( Digits >= Precision ) Precision = Digits+1;
}
}
fclose(F);
/* Compute summary statistics */
sprintf(AvAbsErr, "%.*f", Precision+1, SumAbsErr / N);
sprintf(RelErr, "%.2f", SumAbsErr / SumBaseAbsErr);
sprintf(CorCoeff, "%.2f",
( SumXY - SumX * SumY / N ) /
( qsqrt( (SumXX - SumX * SumX / N) *
(SumYY - SumY * SumY / N) ) + 1E-6 ));
/* If not showing graph, print and quit */
if ( ! Show )
{
printf("Summary:\n--------\n\n");
printf(" Average |error| %10.10s\n", AvAbsErr);
printf(" Relative |error| %4.4s\n", RelErr);
printf(" Correlation coefficient %s\n", CorCoeff);
exit(0);
}
return 0;
}
text.i 0000664 0000764 0000764 00000016620 11501534415 012132 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Text strings for UTF-8 internationalization */
/* ------------------------------------------- */
/* */
/*************************************************************************/
/* General stuff */
//#define UTF8 /* uncomment if using UTF-8 */
#ifdef UTF8
#define CharWidth(S) UTF8CharWidth(S)
#else
#define CharWidth(S) (int) strlen(S)
#endif
/* Strings with width/format restrictions */
/* (W = width when printed) */
#define F_AvErr " Average |error| " /* W=29 */
#define F_RelErr " Relative |error| " /* W=29 */
#define F_CorrCoeff " Correlation coefficient " /* W=29 */
#define F_Actual " Actual" /* W=9 */
#define F_Predicted "Predicted" /* W=9 */
#define F_Value " Value" /* W=9 */
#define F_Case "Case" /* any W */
#define F_UCase "----" /* w same */
/* Strings of arbitrary length */
#define T_Cubist "Cubist"
#define TX_Release(n) "Release " n
#define T_OptHeader "\n Options:\n"
#define T_OptApplication "\tApplication `%s'\n"
#define T_OptMaxRules "\tMaximum of %d rules\n"
#define T_OptExtrap "\tPermit extrapolation of %g%%\n"
#define T_OptUseInst "\tUse instances with rules\n"
#define T_OptAllowInst "\tAllow use of instances with rules\n"
#define TX_NNeighbors(n) "\tUse %d nearest neighbor%s\n", n, Plural(n)
#define T_NoBias "\tUnbiased rules\n"
#define T_OptSampling "\tUse %g%% of data for training\n"
#define T_OptSeed "\tRandom seed %d\n"
#define T_OptXval "\t%d-fold cross-validation\n"
#define T_OptCttee "\t%d-member committee model\n"
#define T_UnregnizedOpt "\n ** Unrecognised option %s %s\n"
#define T_SummaryOpts " ** Summary of options for cubist:\n"
#define T_ListOpts "\t-f \tapplication filestem\n"\
"\t-r \tmaximum number of rules\n"\
"\t-e \tmaximum extrapolation\n"\
"\t-i\t\tuse instances and rules\n"\
"\t-a\t\tallow use of instances and rules\n"\
"\t-n \tnumber of nearest neighbors\n"\
"\t-u\t\tprefer unbiased rules\n"\
"\t-S \ttraining sample percentage\n"\
"\t-I \trandom seed for sampling\n"\
"\t-X \tcross-validate\n"\
"\t-C \tcommittee model\n"\
"\t-h\t\tprint this message\n"
#define T_TargetAtt "\n Target attribute `%s'\n"
#define TX_ReadData(c,a,f) "\nRead %d cases (%d attributes) from"\
" %s.data\n", c, a, f
#define TX_ReadTest(c,f) "Read %d cases from %s.test\n", c, f
#define T_CWtAtt "Using relative case weighting\n"
#define T_AttributesIn "\nAttributes included:\n"
#define T_AttributesOut "\nAttributes excluded:\n"
#define T_EvalTrain "\n\nEvaluation on training data (%d cases%s):\n"
#define T_AttUsage "\n\n\tAttribute usage:\n"\
"\t Conds Model\n\n"
#define T_EvalTest "\nEvaluation on test data (%d cases):\n"
#define T_Time "\n\nTime: %.1f secs\n"
#define T_IgnoreNATarget "\n*** Ignoring cases with N/A target value\n"
#define T_IgnoreBadTarget "\n*** Ignoring cases with unknown or N/A target value\n"
#define T_NoCases "*** No cases with known target values\n"
#define T_ReplaceUnknowns "\n Replacing unknown attribute values:\n"
#define T_NoAppVals "ignoring (no applicable values)"
#define T_By "by"
#define T_SettingNNeighbors "\n\nSetting number of nearest neighbors to %d\n"
#define T_SuggestComposite "Recommend using rules and instances\n"
#define T_SuggestRules "Recommend using rules only\n"
#define T_Rule "Rule"
#define TX_RInfo(c,p,m,l,h,e) ": [%d cases, mean %.*f, range %.7g to %.7g, "\
"est err %.*f]\n\n",c,p,m,l,h,p,e
#define T_If "if"
#define T_Then "then"
#define T_ElementOf "in"
#define T_InRange "in"
#define T_IsUnknown " is unknown\n"
#define T_Default "\n(Default value %.*f)\n\n"
#define T_Summary "Summary"
#define T_FoldsReduced "\n*** folds reduced to number of cases\n"
#define T_Fold "Fold"
#define T_EvalHoldOut "\nEvaluation on hold-out data (%d cases):\n"
#define T_MeanErrMag "Mean |error|"
#define TX_Line(l,f) "*** line %d of `%s':\n ", l, f
#define E_NOFILE(f,e) "cannot open file %s%s\n", f, e
#define E_ForWrite " for writing"
#define E_BADATTNAME "`:' or `:=' expected after attribute name"\
" `%s'\n"
#define E_EOFINATT "unexpected eof while reading attribute `%s'\n"
#define E_SINGLEATTVAL(a,v) "attribute `%s' has only one value `%s'\n",\
a, v
#define E_DUPATTNAME "multiple attributes with name `%s'\n"
#define E_CWTATTERR "case weight attribute must be continuous\n"
#define E_BADATTVAL(v,a) "bad value of `%s' for attribute `%s'\n", v, a
#define E_BADNUMBER(a) "value of `%s' changed to `?'\n", a
#define E_NOMEM "unable to allocate sufficient memory\n"
#define E_TOOMANYVALS(a,n) "too many values for attribute `%s'"\
" (max %d)\n", a, n
#define E_BADDISCRETE "bad number of discrete values for attribute"\
" `%s'\n"
#define E_NOTARGET "target attribute `%s' not found\n"
#define E_BADTARGET "target attribute `%s' is not numeric\n"
#define E_LONGNAME "overlength name: check data file formats\n"
#define E_HITEOF "unexpected end of file\n"
#define E_MISSNAME "missing name or value before `%s'\n"
#define E_BADTSTMP(d,a) "bad timestamp `%s' for attribute `%s'\n", d, a
#define E_BADDATE(d,a) "bad date `%s' for attribute `%s'\n", d, a
#define E_BADTIME(d,a) "bad time `%s' for attribute `%s'\n", d, a
#define E_UNKNOWNATT "unknown attribute name `%s'\n"
#define E_BADDEF1(a,s,x) "in definition of attribute `%s':\n"\
"\tat `%.12s': expect %s\n", a, s, x
#define E_BADDEF2(a,s,x) "in definition of attribute `%s':\n"\
"\t`%s': %s\n", a, s, x
#define E_SAMEATT(a,b) "[warning] attribute `%s' is identical to"\
" attribute `%s'\n", a, b
#define E_BADDEF3 "cannot define target attribute `%s'\n"
#define E_BADDEF4 "[warning] target attribute appears in"\
" definition of attribute `%s':\n"
#define EX_MODELFILE(f) "file %s incompatible with .names file\n", f
#define E_MFATT "undefined or excluded attribute"
#define E_MFATTVAL "undefined attribute value"
#define E_MFEOF "unexpected eof"
#define T_ErrorLimit "\nError limit exceeded\n"
#define TX_IllegalValue(v,l,h) "\t** illegal value %g -- "\
"should be between %g and %g\n", v, l, h
trees.c 0000664 0000764 0000764 00000024053 11501534415 012261 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Routines for displaying, building, saving and restoring trees */
/* ------------------------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define TabSize 4
/* If lines look like getting too long while a tree is being
printed, subtrees are broken off and printed separately after
the main tree is finished */
int SubTree, /* highest subtree to be printed */
SubSpace=0; /* maximum subtree encountered */
Tree *SubDef=Nil; /* pointers to subtrees */
Boolean LastBranch[Width]; /* whether printing last branch of subtree */
/*************************************************************************/
/* */
/* Calculate the depth of nodes in a tree in Errors field */
/* */
/*************************************************************************/
void FindDepth(Tree T)
/* --------- */
{
float MaxDepth=0;
DiscrValue v;
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
FindDepth(T->Branch[v]);
if ( T->Branch[v]->Utility > MaxDepth )
{
MaxDepth = T->Branch[v]->Utility;
}
}
}
T->Utility = MaxDepth + 1;
}
/*************************************************************************/
/* */
/* Display entire decision tree T */
/* */
/*************************************************************************/
void PrintTree(Tree T, String Title)
/* --------- */
{
int s;
FindDepth(T);
SubTree=0;
fprintf(Of, "\n%s\n", Title);
Show(T, 0);
fprintf(Of, "\n");
ForEach(s, 1, SubTree)
{
fprintf(Of, "\nSubTree [S%d]\n", s);
Show(SubDef[s], 0);
fprintf(Of, "\n");
}
}
/*************************************************************************/
/* */
/* Display the tree T with offset Sh */
/* */
/*************************************************************************/
void Show(Tree T, int Sh)
/* ---- */
{
DiscrValue v, MaxV, BrNo=0, Simplest;
Attribute Att;
if ( T->NodeType )
{
/* See whether separate subtree needed */
if ( T != Nil && Sh && Sh * TabSize + MaxLine(T) > Width )
{
if ( ++SubTree >= SubSpace )
{
SubSpace += 100;
SubDef = ( SubDef ? Realloc(SubDef, SubSpace, Tree) :
Alloc(SubSpace, Tree) );
}
SubDef[SubTree] = T;
fprintf(Of, "[S%d]", SubTree);
}
else
{
MaxV = T->Forks;
/* Print simplest branches first */
while ( BrNo < MaxV )
{
Simplest = 1;
ForEach(v, 2, MaxV)
{
if ( T->Branch[v]->Utility < T->Branch[Simplest]->Utility )
{
Simplest = v;
}
}
LastBranch[Sh+1] = ( ++BrNo == MaxV );
ShowBranch(Sh, T, Simplest, BrNo);
T->Branch[Simplest]->Utility = 1E10;
}
}
}
else
{
fprintf(Of, " AV %g (%d:%g)", T->Mean, T->Cases, T->Coeffs);
fprintf(Of, " [%g", T->Model[0]);
ForEach(Att, 1, MaxAtt)
{
if ( T->Model[Att] )
{
fprintf(Of, " + %g %s", T->Model[Att], AttName[Att]);
}
}
fprintf(Of, "]");
}
}
/*************************************************************************/
/* */
/* Print a node T with offset Sh, branch value v, and continue */
/* */
/*************************************************************************/
void ShowBranch(int Sh, Tree T, DiscrValue v, DiscrValue BrNo)
/* ---------- */
{
DiscrValue Pv, Last;
Attribute Att;
Boolean FirstValue;
int TextWidth, Skip, Values=0, i, Extra;
Att = T->Tested;
switch ( T->NodeType )
{
case BrDiscr:
Indent(Sh, BrNo);
fprintf(Of, "%s = %s:", AttName[Att], AttValName[Att][v]);
break;
case BrThresh:
Indent(Sh, BrNo);
if ( v == 1 )
{
fprintf(Of, "%s = N/A:", AttName[Att]);
}
else
{
fprintf(Of, "%s %s %.*g:",
AttName[Att], ( v == 2 ? "<=" : ">" ), PREC, T->Cut);
}
break;
case BrSubset:
/* Count values at this branch */
ForEach(Pv, 1, MaxAttVal[Att])
{
if ( In(Pv, T->Subset[v]) )
{
Last = Pv;
Values++;
}
}
if ( ! Values ) return;
Indent(Sh, BrNo);
if ( Values == 1 )
{
fprintf(Of, "%s = %s:", AttName[Att], AttValName[Att][Last]);
break;
}
fprintf(Of, "%s in {", AttName[Att]);
FirstValue = true;
Skip = strlen(AttName[Att]) + 5;
TextWidth = Skip + Sh * TabSize;
ForEach(Pv, 1, Last)
{
if ( In(Pv, T->Subset[v]) )
{
Extra = ( Pv != Last || T->Branch[v]->NodeType ? 0 :
6 );
if ( ! FirstValue &&
TextWidth + strlen(AttValName[Att][Pv]) + 11 + Extra > Width )
{
Indent(Sh, 0);
ForEach(i, 1, Skip) putc(' ', Of);
TextWidth = Skip + Sh * TabSize;
FirstValue = true;
}
fprintf(Of, "%s%c",
AttValName[Att][Pv], Pv == Last ? '}' : ',');
TextWidth += strlen(AttValName[Att][Pv]) + 1;
FirstValue = false;
}
}
putc(':', Of);
}
Show(T->Branch[v], Sh+1);
}
/*************************************************************************/
/* */
/* Find the maximum single line size for non-leaf subtree T */
/* */
/*************************************************************************/
int MaxLine(Tree T)
/* ------- */
{
Attribute Att;
DiscrValue v, vv;
int Ll, One, MaxLl=0;
Att = T->Tested;
/* First find the max length of the line excluding tested att */
ForEach(v, 1, T->Forks)
{
switch ( T->NodeType )
{
case BrThresh:
Ll = 4;
break;
case BrDiscr:
Ll = strlen(AttValName[Att][v]) + 1;
break;
case BrSubset: /* difficult! */
Ll = 0;
ForEach(vv, 1, MaxAttVal[Att])
{
if ( In(vv,T->Subset[v]) )
{
One = strlen(AttValName[Att][vv]) + 6;
if ( One > Ll ) Ll = One;
}
}
}
/* Check whether ends in leaf */
if ( ! T->Branch[v]->NodeType )
{
Ll += 6;
}
if ( Ll > MaxLl ) MaxLl = Ll;
}
return strlen(AttName[Att]) + 4 + MaxLl;
}
/*************************************************************************/
/* */
/* Indent Sh columns */
/* */
/*************************************************************************/
void Indent(int Sh, int BrNo)
/* ------ */
{
int i;
fprintf(Of, "\n");
for ( i = 1 ; i <= Sh ; i++ )
{
fprintf(Of, "%s", ( i == Sh && BrNo == 1 ? ":..." :
LastBranch[i] ? " " : ": " ));
}
}
/*************************************************************************/
/* */
/* Free up space taken up by tree Node */
/* */
/*************************************************************************/
void FreeTree(Tree T)
/* -------- */
{
DiscrValue v;
if ( ! T ) return;
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
FreeTree(T->Branch[v]);
}
Free(T->Branch); T->Branch = Nil;
if ( T->NodeType == BrSubset )
{
FreeVector((void **) T->Subset, 1, T->Forks); T->Subset = Nil;
}
}
FreeUnlessNil(T->Model); T->Model = Nil;
FreeUnlessNil(T->MCopy); T->MCopy = Nil;
Free(T);
}
/*************************************************************************/
/* */
/* Construct a leaf in a given node */
/* */
/*************************************************************************/
Tree Leaf(CaseCount Cases, double Mean, double SD)
/* ---- */
{
Tree Node;
Node = AllocZero(1, TreeRec);
Node->NodeType = 0;
Node->Cases = Cases;
Node->Mean = Mean;
Node->SD = SD;
Node->Model = AllocZero(MaxAtt+1, double);
Node->Model[0] = Mean;
return Node;
}
/*************************************************************************/
/* */
/* Insert branches in a node */
/* */
/*************************************************************************/
void Sprout(Tree T, DiscrValue Branches)
/* ------ */
{
T->Forks = Branches;
T->Branch = Alloc(Branches+1, Tree);
}
/*************************************************************************/
/* */
/* Count the nodes in a tree */
/* */
/*************************************************************************/
int TreeSize(Tree T)
/* -------- */
{
int Sum=1;
DiscrValue v;
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
Sum += TreeSize(T->Branch[v]);
}
return Sum;
}
return 1;
}
/*************************************************************************/
/* */
/* Count the non-null leaves in a tree */
/* */
/*************************************************************************/
int TreeLeaves(Tree T)
/* ---------- */
{
int Sum=0;
DiscrValue v;
if ( ! T || T->Cases < 1 ) return 0;
if ( T->NodeType )
{
ForEach(v, 1, T->Forks)
{
Sum += TreeLeaves(T->Branch[v]);
}
return Sum;
}
return 1;
}
update.c 0000664 0000764 0000764 00000007466 11501534415 012432 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Routines that provide information on progress */
/* --------------------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
int Stage=0; /* Current stage number */
FILE *Uf=0; /* File to which update info written */
/*************************************************************************/
/* */
/* There are twelve stages (see messages in Progress() below) */
/* Record stage and open update file if necessary */
/* */
/*************************************************************************/
void NotifyStage(int S)
/* ----------- */
{
Stage = S;
if ( S == 1 )
{
if ( ! (Uf = GetFile(".tmp", "w")) ) Error(0, "", E_ForWrite);
}
}
/*************************************************************************/
/* */
/* Print progress message. This routine is called in two ways: */
/* * negative Delta = measure of total effort required for stage */
/* * positive Delta = increment since last call */
/* */
/*************************************************************************/
void Progress(float Delta)
/* -------- */
{
static float Total, Current=0;
static int Twentieth, Percent=-6;
int p;
static char *Message[]={ "",
"Reading training data ",
"Grouping training cases ",
"Adding linear models ",
"Simplifying groups ",
"Forming rules ",
"Preparing instance index ",
"Setting number of neighbors",
"Assessing composite models ",
"Evaluating on training data",
"Reading test data ",
"Evaluating on test data ",
"Cleaning up " },
Tell[]={ 0,0,1,1,1,1,0,1,1,1,0,1,0,0 },
*Unit[]={ "",
"",
"cases covered",
"models added",
"groups discarded",
"cases covered",
"",
"cases assessed",
"cases assessed",
"cases checked",
"",
"cases checked",
"" },
*Done=">>>>>>>>>>>>>>>>>>>>",
*ToDo="....................";
if ( Delta < 0 )
{
Total = -Delta;
Current = 0;
Percent = -6;
}
else
{
Current += Delta;
}
if ( (p = (100 * Current) / Total) == 100 || p >= Percent + 5 )
{
Percent = p;
Twentieth = p / 5;
fprintf(Uf, "%s", Message[Stage]);
if ( Tell[Stage] )
{
fprintf(Uf, " %s%s (%d %s)",
Done + (20 - Twentieth), ToDo + Twentieth,
(int) (Current+0.5),
Unit[Stage]);
}
fprintf(Uf, "\n");
fflush(Uf);
}
}
utility.c 0000664 0000764 0000764 00000055760 11501534415 012653 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Print header for Cubist programs */
/* -------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
#define NAME T_Cubist
void PrintHeader(String Title)
/* ----------- */
{
char TitleLine[80];
time_t clock;
int Underline;
clock = time(0);
sprintf(TitleLine, "%s%s [%s]", NAME, Title, TX_Release(RELEASE));
fprintf(Of, "\n%s %s", TitleLine, ctime(&clock));
Underline = CharWidth(TitleLine);
while ( Underline-- ) putc('-', Of);
putc('\n', Of);
}
/*************************************************************************/
/* */
/* This is a specialised form of the getopt utility. */
/* */
/*************************************************************************/
String OptArg, Option;
char ProcessOption(int Argc, char *Argv[], char *Options)
/* ------------- */
{
int i;
static int OptNo=1;
if ( OptNo >= Argc ) return '\00';
if ( *(Option = Argv[OptNo++]) != '-' ) return '?';
for ( i = 0 ; Options[i] ; i++ )
{
if ( Options[i] == Option[1] )
{
OptArg = ( Options[i+1] != '+' ? Nil :
Option[2] ? Option+2 :
OptNo < Argc ? Argv[OptNo++] : "0" );
return Option[1];
}
}
return '?';
}
/*************************************************************************/
/* */
/* Protected memory allocation routines */
/* */
/*************************************************************************/
void *Pmalloc(size_t Bytes)
/* ------- */
{
void *p=Nil;
if ( ! Bytes || (p = (void *) malloc(Bytes)) )
{
return p;
}
Error(NOMEM, "", "");
return Nil; /* to keep compilers happy */
}
void *Prealloc(void *Present, size_t Bytes)
/* -------- */
{
void *p=Nil;
if ( ! Bytes ) return Nil;
if ( ! Present ) return Pmalloc(Bytes);
if ( (p = (void *) realloc(Present, Bytes)) )
{
return p;
}
Error(NOMEM, "", "");
return Nil; /* to keep compilers happy */
}
void *Pcalloc(size_t Number, unsigned int Size)
/* ------- */
{
void *p=Nil;
if ( ! Number || (p = (void *) calloc(Number, Size)) )
{
return p;
}
Error(NOMEM, "", "");
return Nil; /* to keep compilers happy */
}
void FreeVector(void **V, int First, int Last)
/* ---------- */
{
if ( V )
{
while ( First <= Last )
{
FreeUnlessNil(V[First]);
First++;
}
Free(V);
}
}
/*************************************************************************/
/* */
/* Special memory allocation routines for case memory */
/* */
/*************************************************************************/
typedef struct _datablockrec *DataBlock;
typedef struct _datablockrec
{
DataRec Head; /* first address */
int Allocated; /* number of cases in this block */
DataBlock Prev; /* previous data block */
}
DataBlockRec;
DataBlock DataMem=Nil;
int DataBlockSize=0;
DataRec NewCase(void)
/* ------- */
{
DataBlock Prev;
if ( ! DataMem || DataMem->Allocated == DataBlockSize )
{
DataBlockSize = Min(8192, 262144 / (MaxAtt+3) + 1);
Prev = DataMem;
DataMem = AllocZero(1, DataBlockRec);
DataMem->Head = Alloc(DataBlockSize * (MaxAtt+3), AttValue);
DataMem->Prev = Prev;
}
return DataMem->Head + (DataMem->Allocated++) * (MaxAtt+3);
}
void FreeCases(void)
/* --------- */
{
DataBlock Prev;
while ( DataMem )
{
Prev = DataMem->Prev;
Free(DataMem->Head);
Free(DataMem);
DataMem = Prev;
}
}
void FreeLastCase(DataRec Case)
/* ------------ */
{
DataMem->Allocated--;
}
/*************************************************************************/
/* */
/* Generate uniform random numbers */
/* */
/*************************************************************************/
#define Modify(F,S) if ( (F -= S) < 0 ) F += 1.0
int KRFp=0, KRSp=0;
double KRandom(void)
/* ------- */
{
static double URD[55];
double V1, V2;
int i, j;
/* Initialisation */
if ( KRFp == KRSp )
{
KRFp = 0;
KRSp = 31;
V1 = 1.0;
V2 = 0.314159285;
ForEach(i, 1, 55)
{
URD[ j = (i * 21) % 55 ] = V1;
V1 = V2 - V1;
if ( V1 < 0 ) V1 += 1.0;
V2 = URD[j];
}
ForEach(j, 0, 5)
{
ForEach(i, 0, 54)
{
Modify(URD[i], URD[(i+30) % 55]);
}
}
}
KRFp = (KRFp + 1) % 55;
KRSp = (KRSp + 1) % 55;
Modify(URD[KRFp], URD[KRSp]);
return URD[KRFp];
}
void ResetKR(int KRInit)
/* ------- */
{
KRFp = KRSp = 0;
KRInit += 1000;
while ( KRInit-- )
{
KRandom();
}
}
/*************************************************************************/
/* */
/* Error messages */
/* */
/*************************************************************************/
void Error(int ErrNo, String S1, String S2)
/* ----- */
{
Boolean Quit=false, WarningOnly=false;
char Buffer[10000], *Msg=Buffer;
if ( ! ErrMsgs ) fprintf(Of, "\n");
if ( ErrNo == NOFILE || ErrNo == NOMEM || ErrNo == MODELFILE )
{
sprintf(Msg, "*** ");
}
else
{
sprintf(Msg, TX_Line(LineNo, Fn));
}
Msg += strlen(Buffer);
switch ( ErrNo )
{
case NOFILE:
sprintf(Msg, E_NOFILE(Fn, S2));
Quit = true;
break;
case BADATTNAME:
sprintf(Msg, E_BADATTNAME, S1);
break;
case EOFINATT:
sprintf(Msg, E_EOFINATT, S1);
break;
case SINGLEATTVAL:
sprintf(Msg, E_SINGLEATTVAL(S1, S2));
break;
case DUPATTNAME:
sprintf(Msg, E_DUPATTNAME, S1);
break;
case CWTATTERR:
sprintf(Msg, E_CWTATTERR);
break;
case BADATTVAL:
sprintf(Msg, E_BADATTVAL(S2, S1));
break;
case BADNUMBER:
sprintf(Msg, E_BADNUMBER(S1));
break;
case NOMEM:
sprintf(Msg, E_NOMEM);
Quit = true;
break;
case TOOMANYVALS:
sprintf(Msg, E_TOOMANYVALS(S1, (int) (long) S2));
Quit = true;
break;
case BADDISCRETE:
sprintf(Msg, E_BADDISCRETE, S1);
Quit = true;
break;
case NOTARGET:
sprintf(Msg, E_NOTARGET, S1);
Quit = true;
break;
case BADTARGET:
sprintf(Msg, E_BADTARGET, S1);
Quit = true;
break;
case LONGNAME:
sprintf(Msg, E_LONGNAME);
Quit = true;
break;
case HITEOF:
sprintf(Msg, E_HITEOF);
break;
case MISSNAME:
sprintf(Msg, E_MISSNAME, S2);
break;
case BADTSTMP:
sprintf(Msg, E_BADTSTMP(S2, S1));
break;
case BADDATE:
sprintf(Msg, E_BADDATE(S2, S1));
break;
case BADTIME:
sprintf(Msg, E_BADTIME(S2, S1));
break;
case UNKNOWNATT:
sprintf(Msg, E_UNKNOWNATT, S1);
break;
case BADDEF1:
sprintf(Msg, E_BADDEF1(AttName[MaxAtt], S1, S2));
break;
case BADDEF2:
sprintf(Msg, E_BADDEF2(AttName[MaxAtt], S1, S2));
break;
case SAMEATT:
sprintf(Msg, E_SAMEATT(AttName[MaxAtt], S1));
WarningOnly = true;
break;
case BADDEF3:
sprintf(Msg, E_BADDEF3, AttName[MaxAtt]);
break;
case BADDEF4:
sprintf(Msg, E_BADDEF4, AttName[MaxAtt]);
WarningOnly = true;
break;
case MODELFILE:
sprintf(Msg, EX_MODELFILE(Fn));
sprintf(Msg, " (%s `%s')\n", S1, S2);
Quit = true;
break;
}
fprintf(Of, Buffer);
if ( ! WarningOnly ) ErrMsgs++;
if ( ErrMsgs == 10 )
{
fprintf(Of, T_ErrorLimit);
MaxCase--;
Quit = true;
}
if ( Quit && Of )
{
Goodbye(1);
}
}
/*************************************************************************/
/* */
/* Generate the label for a case */
/* */
/*************************************************************************/
char LabelBuffer[1000];
String CaseLabel(CaseNo N)
/* --------- */
{
String p;
if ( LabelAtt && (p = IgnoredVals + SVal(Case[N], LabelAtt)) )
;
else
{
sprintf(LabelBuffer, "#%d", N+1);
p = LabelBuffer;
}
return p;
}
/*************************************************************************/
/* */
/* Open file with given extension for read/write */
/* */
/*************************************************************************/
FILE *GetFile(String Extension, String RW)
/* -------- */
{
strcpy(Fn, FileStem);
strcat(Fn, Extension);
return fopen(Fn, RW);
}
/*************************************************************************/
/* */
/* Determine total elapsed time so far. */
/* */
/*************************************************************************/
#include
double ExecTime(void)
/* -------- */
{
struct timeval TV;
struct timezone TZ={0,0};
gettimeofday(&TV, &TZ);
return TV.tv_sec + TV.tv_usec / 1000000.0;
}
/*************************************************************************/
/* */
/* Determine precision of floating value (max 6 digits) */
/* */
/*************************************************************************/
int Denominator(ContValue Val)
/* ----------- */
{
double Resid, IntPart;
int Mult=1, MaxMult=1000000;
Resid = fabs(Val);
if ( Resid == 0 ) return 1;
/* Skip leading zeros */
while ( Resid < 0.1 )
{
Resid *= 10;
Mult *= 10;
}
Resid = modf(Resid, &IntPart);
/* Reduce maximum denominator by integral digits */
while ( IntPart >= 1 && MaxMult >= 10 )
{
IntPart /= 10;
MaxMult /= 10;
}
for ( ; Mult < MaxMult ; Mult *= 10 )
{
if ( Resid < 0.005 || Resid > 0.995 ) return Mult;
Resid = modf(Resid * 10, &IntPart);
}
return MaxMult;
}
int FracBase(Attribute Att)
/* -------- */
{
CaseNo i;
int Denom=0, ThisDenom;
ForEach(i, 0, MaxCase)
{
if ( CVal(Case[i], Att) != UNKNOWN &&
! NotApplic(Case[i], Att) &&
(ThisDenom = Denominator(CVal(Case[i], Att))) > Denom )
{
Denom = ThisDenom;
if ( Denom == 1000000 ) break;
}
}
return Denom;
}
/*************************************************************************/
/* */
/* Routines to process date (Algorithm due to Gauss?) */
/* */
/*************************************************************************/
int GetInt(String S, int N)
/* ------ */
{
int Result=0;
while ( N-- )
{
if ( ! isdigit(*S) ) return 0;
Result = Result * 10 + (*S++ - '0');
}
return Result;
}
int DateToDay(String DS) /* Day 1 is 0000/03/01 */
/* --------- */
{
int Year, Month, Day;
if ( strlen(DS) != 10 ) return 0;
Year = GetInt(DS, 4);
Month = GetInt(DS+5, 2);
Day = GetInt(DS+8, 2);
if ( ! ( DS[4] == '/' && DS[7] == '/' || DS[4] == '-' && DS[7] == '-' ) ||
Year < 0 || Month < 1 || Day < 1 ||
Month > 12 ||
Day > 31 ||
Day > 30 &&
( Month == 4 || Month == 6 || Month == 9 || Month == 11 ) ||
Month == 2 &&
( Day > 29 ||
Day > 28 && ( Year % 4 != 0 ||
Year % 100 == 0 && Year % 400 != 0 ) ) )
{
return 0;
}
if ( (Month -= 2) <= 0 )
{
Month += 12;
Year -= 1;
}
return Year * 365 + Year / 4 - Year / 100 + Year / 400
+ 367 * Month / 12
+ Day - 30;
}
void DayToDate(int Day, String Date)
/* --------- */
{
int Year, Month, OrigDay=Day;
if ( Day <= 0 )
{
strcpy(Date, "?");
return;
}
Year = (Day - 1) / 365.2425L; /* Year = completed years */
Day -= Year * 365 + Year / 4 - Year / 100 + Year / 400;
if ( Day < 1 )
{
Year--;
Day = OrigDay - (Year * 365 + Year / 4 - Year / 100 + Year / 400);
}
else
if ( Day > 366 ||
Day == 366 &&
( (Year+1) % 4 != 0 || (Year+1) % 100 == 0 && (Year+1) % 400 != 0 ) )
{
Year++;
Day = OrigDay - (Year * 365 + Year / 4 - Year / 100 + Year / 400);
}
Month = (Day + 30) * 12 / 367;
Day -= 367 * Month / 12 - 30;
if ( Day < 1 )
{
Month = 11;
Day = 31;
}
Month += 2;
if ( Month > 12 )
{
Month -= 12;
Year++;
}
sprintf(Date, "%d/%d%d/%d%d", Year, Month/10, Month % 10, Day/10, Day % 10);
}
/*************************************************************************/
/* */
/* Routines to process clock time and timestamps */
/* */
/*************************************************************************/
int TimeToSecs(String TS)
/* ---------- */
{
int Hour, Mins, Secs;
if ( strlen(TS) != 8 ) return -1;
Hour = GetInt(TS, 2);
Mins = GetInt(TS+3, 2);
Secs = GetInt(TS+6, 2);
if ( TS[2] != ':' || TS[5] != ':' ||
Hour >= 24 || Mins >= 60 || Secs >= 60 )
{
return -1;
}
return Hour * 3600 + Mins * 60 + Secs;
}
void SecsToTime(int Secs, String Time)
/* ---------- */
{
int Hour, Mins;
Hour = Secs / 3600;
Mins = (Secs % 3600) / 60;
Secs = Secs % 60;
sprintf(Time, "%d%d:%d%d:%d%d",
Hour / 10, Hour % 10,
Mins / 10, Mins % 10,
Secs / 10, Secs % 10);
}
void SetTSBase(int y)
/* --------- */
{
y -= 15;
TSBase = y * 365 + y / 4 - y / 100 + y / 400 + (367 * 4) / 12 + 1 - 30;
}
int TStampToMins(String TS)
/* ------------ */
{
int Day, Sec, i;
/* Check for reasonable length and space between date and time */
if ( strlen(TS) < 19 || ! Space(TS[10]) ) return (1 << 30);
/* Read date part */
TS[10] = '\00';
Day = DateToDay(TS);
TS[10] = ' ';
/* Skip one or more spaces */
for ( i = 11 ; TS[i] && Space(TS[i]) ; i++ )
;
/* Read time part */
Sec = TimeToSecs(TS+i);
/* Return a long time in the future if there is an error */
return ( Day < 1 || Sec < 0 ? (1 << 30) :
(Day - TSBase) * 1440 + (Sec + 30) / 60 );
}
/*************************************************************************/
/* */
/* Convert a continuous value to a string. DS must be large */
/* enough to hold any value (e.g. a date, time, ...) */
/* */
/*************************************************************************/
void CValToStr(ContValue CV, Attribute Att, String DS)
/* --------- */
{
int Mins;
if ( TStampVal(Att) )
{
DayToDate(floor(CV / 1440) + TSBase, DS);
DS[10] = ' ';
Mins = rint(CV) - floor(CV / 1440) * 1440;
SecsToTime(Mins * 60, DS+11);
}
else
if ( DateVal(Att) )
{
DayToDate(CV, DS);
}
else
if ( TimeVal(Att) )
{
SecsToTime(CV, DS);
}
else
{
sprintf(DS, "%.*g", PREC, CV);
}
}
void Check(float Val, float Low, float High)
/* ----- */
{
if ( Val < Low || Val > High )
{
fprintf(Of, TX_IllegalValue(Val, Low, High));
exit(1);
}
}
/*************************************************************************/
/* */
/* Deallocate all dynamic storage */
/* */
/*************************************************************************/
void Cleanup(void)
/* ------- */
{
RuleNo r;
extern FILE *Uf;
NotifyStage(CLEANUP);
FreeEnvData();
CheckClose(Mf); Mf = 0;
CheckClose(Pf); Pf = 0;
CheckClose(Uf); Uf = 0;
FreeFormRuleData();
/* If cross-val, restore MaxCase */
if ( Blocked )
{
MaxCase = SaveMaxCase;
Free(Blocked); Blocked = Nil;
}
if ( Instance )
{
FreeInstances(); Instance = Nil;
}
if ( MaxCase >= 0 )
{
FreeData(Case);
Case = Nil;
}
FreeUnlessNil(SaveCase); SaveCase = Nil;
if ( Cttee )
{
FreeCttee(Cttee); Cttee = Nil;
}
/* Release intermediate tree/rules if interrupted */
if ( TempMT )
{
FreeTree(TempMT); TempMT = Nil;
}
if ( Rule )
{
ForEach(r, 1, NRules)
{
ReleaseRule(Rule[r]);
}
Free(Rule); Rule = Nil;
}
KDTree = Nil;
FreeNamesData();
MaxCase = -1;
NotifyStage(0);
}
#ifdef UTF8
///////////////////////////////////////////////////////////////////////////
// //
// Routines for Unicode/UTF-8 processing //
// ------------------------------------- //
// //
///////////////////////////////////////////////////////////////////////////
#include
/*************************************************************************/
/* */
/* Determine the total character width of a UTF-8 string */
/* */
/*************************************************************************/
int UTF8CharWidth(unsigned char *U)
/* ------------- */
{
int CWidth=0, Mask, This;
wchar_t Unicode;
while ( *U )
{
Unicode = *U;
if ( *U < 0x7F )
{
/* ASCII character */
CWidth++;
U++;
}
else
{
/* Discard header bits */
Mask = 0x80;
while ( Unicode & Mask )
{
Unicode ^= Mask;
Mask = Mask >> 1;
}
while ( ((*(++U)) & 0xc0) == 0x80 )
{
Unicode = (Unicode << 6) | (*U & 0x3f);
}
if ( (This = wcwidth(Unicode)) > 0 ) CWidth += This;
}
}
return CWidth;
}
////////////////////////////////////////////////////////////////////////////////
// Public domain code to determine the width of a Unicode character //
////////////////////////////////////////////////////////////////////////////////
/*
* This is an implementation of wcwidth() and wcswidth() as defined in
* "The Single UNIX Specification, Version 2, The Open Group, 1997"
*
*
* Markus Kuhn -- 2000-02-08 -- public domain
*/
//#include
/* These functions define the column width of an ISO 10646 character
* as follows:
*
* - The null character (U+0000) has a column width of 0.
*
* - Other C0/C1 control characters and DEL will lead to a return
* value of -1.
*
* - Non-spacing and enclosing combining characters (general
* category code Mn or Me in the Unicode database) have a
* column width of 0.
*
* - Spacing characters in the East Asian Wide (W) or East Asian
* FullWidth (F) category as defined in Unicode Technical
* Report #11 have a column width of 2.
*
* - All remaining characters (including all printable
* ISO 8859-1 and WGL4 characters, Unicode control characters,
* etc.) have a column width of 1.
*
* This implementation assumes that wchar_t characters are encoded
* in ISO 10646.
*/
int wcwidth(wchar_t ucs)
{
/* sorted list of non-overlapping intervals of non-spacing characters */
static const struct interval {
unsigned short first;
unsigned short last;
} combining[] = {
{ 0x0300, 0x034E }, { 0x0360, 0x0362 }, { 0x0483, 0x0486 },
{ 0x0488, 0x0489 }, { 0x0591, 0x05A1 }, { 0x05A3, 0x05B9 },
{ 0x05BB, 0x05BD }, { 0x05BF, 0x05BF }, { 0x05C1, 0x05C2 },
{ 0x05C4, 0x05C4 }, { 0x064B, 0x0655 }, { 0x0670, 0x0670 },
{ 0x06D6, 0x06E4 }, { 0x06E7, 0x06E8 }, { 0x06EA, 0x06ED },
{ 0x0711, 0x0711 }, { 0x0730, 0x074A }, { 0x07A6, 0x07B0 },
{ 0x0901, 0x0902 }, { 0x093C, 0x093C }, { 0x0941, 0x0948 },
{ 0x094D, 0x094D }, { 0x0951, 0x0954 }, { 0x0962, 0x0963 },
{ 0x0981, 0x0981 }, { 0x09BC, 0x09BC }, { 0x09C1, 0x09C4 },
{ 0x09CD, 0x09CD }, { 0x09E2, 0x09E3 }, { 0x0A02, 0x0A02 },
{ 0x0A3C, 0x0A3C }, { 0x0A41, 0x0A42 }, { 0x0A47, 0x0A48 },
{ 0x0A4B, 0x0A4D }, { 0x0A70, 0x0A71 }, { 0x0A81, 0x0A82 },
{ 0x0ABC, 0x0ABC }, { 0x0AC1, 0x0AC5 }, { 0x0AC7, 0x0AC8 },
{ 0x0ACD, 0x0ACD }, { 0x0B01, 0x0B01 }, { 0x0B3C, 0x0B3C },
{ 0x0B3F, 0x0B3F }, { 0x0B41, 0x0B43 }, { 0x0B4D, 0x0B4D },
{ 0x0B56, 0x0B56 }, { 0x0B82, 0x0B82 }, { 0x0BC0, 0x0BC0 },
{ 0x0BCD, 0x0BCD }, { 0x0C3E, 0x0C40 }, { 0x0C46, 0x0C48 },
{ 0x0C4A, 0x0C4D }, { 0x0C55, 0x0C56 }, { 0x0CBF, 0x0CBF },
{ 0x0CC6, 0x0CC6 }, { 0x0CCC, 0x0CCD }, { 0x0D41, 0x0D43 },
{ 0x0D4D, 0x0D4D }, { 0x0DCA, 0x0DCA }, { 0x0DD2, 0x0DD4 },
{ 0x0DD6, 0x0DD6 }, { 0x0E31, 0x0E31 }, { 0x0E34, 0x0E3A },
{ 0x0E47, 0x0E4E }, { 0x0EB1, 0x0EB1 }, { 0x0EB4, 0x0EB9 },
{ 0x0EBB, 0x0EBC }, { 0x0EC8, 0x0ECD }, { 0x0F18, 0x0F19 },
{ 0x0F35, 0x0F35 }, { 0x0F37, 0x0F37 }, { 0x0F39, 0x0F39 },
{ 0x0F71, 0x0F7E }, { 0x0F80, 0x0F84 }, { 0x0F86, 0x0F87 },
{ 0x0F90, 0x0F97 }, { 0x0F99, 0x0FBC }, { 0x0FC6, 0x0FC6 },
{ 0x102D, 0x1030 }, { 0x1032, 0x1032 }, { 0x1036, 0x1037 },
{ 0x1039, 0x1039 }, { 0x1058, 0x1059 }, { 0x17B7, 0x17BD },
{ 0x17C6, 0x17C6 }, { 0x17C9, 0x17D3 }, { 0x18A9, 0x18A9 },
{ 0x20D0, 0x20E3 }, { 0x302A, 0x302F }, { 0x3099, 0x309A },
{ 0xFB1E, 0xFB1E }, { 0xFE20, 0xFE23 }
};
int min = 0;
int max = sizeof(combining) / sizeof(struct interval) - 1;
int mid;
/* test for 8-bit control characters */
if (ucs == 0)
return 0;
if (ucs < 32 || (ucs >= 0x7f && ucs < 0xa0))
return -1;
/* first quick check for Latin-1 etc. characters */
if (ucs < combining[0].first)
return 1;
/* binary search in table of non-spacing characters */
while (max >= min) {
mid = (min + max) / 2;
if (combining[mid].last < ucs)
min = mid + 1;
else if (combining[mid].first > ucs)
max = mid - 1;
else if (combining[mid].first <= ucs && combining[mid].last >= ucs)
return 0;
}
/* if we arrive here, ucs is not a combining or C0/C1 control character */
/* fast test for majority of non-wide scripts */
if (ucs < 0x1100)
return 1;
return 1 +
((ucs >= 0x1100 && ucs <= 0x115f) || /* Hangul Jamo */
(ucs >= 0x2e80 && ucs <= 0xa4cf && (ucs & ~0x0011) != 0x300a &&
ucs != 0x303f) || /* CJK ... Yi */
(ucs >= 0xac00 && ucs <= 0xd7a3) || /* Hangul Syllables */
(ucs >= 0xf900 && ucs <= 0xfaff) || /* CJK Compatibility Ideographs */
(ucs >= 0xfe30 && ucs <= 0xfe6f) || /* CJK Compatibility Forms */
(ucs >= 0xff00 && ucs <= 0xff5f) || /* Fullwidth Forms */
(ucs >= 0xffe0 && ucs <= 0xffe6));
}
int wcswidth(const wchar_t *pwcs, size_t n)
{
int w, width = 0;
for (;*pwcs && n-- > 0; pwcs++)
if ((w = wcwidth(*pwcs)) < 0)
return -1;
else
width += w;
return width;
}
#endif
xval.c 0000664 0000764 0000764 00000015110 11501534415 012103 0 ustar quinlan quinlan /*************************************************************************/
/* */
/* Copyright 2010 Rulequest Research Pty Ltd. */
/* */
/* This file is part of Cubist GPL Edition, a single-threaded version */
/* of Cubist release 2.07. */
/* */
/* Cubist GPL Edition is free software: you can redistribute it and/or */
/* modify it under the terms of the GNU General Public License as */
/* published by the Free Software Foundation, either version 3 of the */
/* License, or (at your option) any later version. */
/* */
/* Cubist GPL Edition is distributed in the hope that it will be */
/* useful, but WITHOUT ANY WARRANTY; without even the implied warranty */
/* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */
/* GNU General Public License for more details. */
/* */
/* You should have received a copy of the GNU General Public License */
/* (gpl.txt) along with Cubist GPL Edition. If not, see */
/* */
/* . */
/* */
/*************************************************************************/
/*************************************************************************/
/* */
/* Carry out cross validation trials */
/* --------------------------------- */
/* */
/*************************************************************************/
#include "defns.i"
#include "extern.i"
/*************************************************************************/
/* */
/* Outer function (differs from xval script) */
/* */
/*************************************************************************/
void CrossVal(void)
/* -------- */
{
CaseNo i, Size, Start=0, Next, N;
int f, SmallTestBlocks;
double *ErrMag,
R, SumR=0, SumRR=0, /* real value */
P, SumP=0, SumPP=0, /* predicted value */
SumRP=0, /* for correlation coeff */
M, SumM=0, /* error magnitude */
D, SumD=0, /* default error magnitude */
TrainingMean;
if ( FOLDS > MaxCase+1 )
{
fprintf(Of, T_FoldsReduced);
FOLDS = MaxCase+1;
}
ErrMag = AllocZero(FOLDS, double);
Blocked = Alloc(MaxCase+1, DataRec);
SaveMaxCase = MaxCase;
Prepare();
if ( ! (Pf = GetFile(".pred", "w")) ) Error(0, Fn, E_ForWrite);
/* First test blocks may be smaller than the others */
SmallTestBlocks = FOLDS - ((MaxCase+1) % FOLDS);
Size = (MaxCase+1) / FOLDS;
ForEach(f, 0, FOLDS-1)
{
fprintf(Of, "\n\n[ " T_Fold " %d ]\n", f+1);
if ( f == SmallTestBlocks ) Size++;
MaxCase = SaveMaxCase - Size;
/* Copy blocks back into Case:
train: 0 to MaxCase
test: MaxCase+1 to SaveMaxCase */
Next = Start;
ForEach(i, 0, SaveMaxCase)
{
Case[i] = Blocked[Next];
Next = (Next + 1) % (SaveMaxCase + 1);
}
Start = (Start + MaxCase + 1) % (SaveMaxCase + 1);
TrainingMean = 0;
ForEach(i, 0, MaxCase)
{
TrainingMean += Class(Case[i]);
}
TrainingMean /= MaxCase+1;
ConstructCttee();
/* Record size and errors */
FindPredictedValues(Cttee, MaxCase+1, SaveMaxCase);
fprintf(Pf, "\n(Default value %.*f)\n\n", Precision+1, GlobalMean);
fprintf(Pf, " Actual Predicted Case\n"
" Value Value\n"
" -------- --------- ----\n");
ForEach(i, MaxCase+1, SaveMaxCase)
{
R = Class(Case[i]);
P = PredVal(Case[i]);
ErrMag[f] += (M = fabs(Class(Case[i]) - P));
D = fabs(R - TrainingMean);
fprintf(Pf, "%9.*f %9.*f %s\n",
Precision, R, Precision+1, P, CaseLabel(i));
/* Statistics for PVE, correlation coefficient */
SumM += M;
SumR += R;
SumRR += R * R;
SumP += P;
SumPP += P * P;
SumRP += R * P;
SumD += D;
}
ErrMag[f] /= Size;
fprintf(Of, T_EvalHoldOut "\n " T_MeanErrMag " %.*f\n",
Size, Precision+1, ErrMag[f]);
/* Free space used by rulesets */
FreeCttee(Cttee); Cttee = Nil;
if ( USEINSTANCES )
{
FreeInstances();
}
}
fclose(Pf);
Pf = 0;
/* Print summary of cross-validation */
MaxCase = SaveMaxCase;
N = MaxCase+1;
fprintf(Of, "\n\n" T_Summary ":\n\n");
fprintf(Of, F_AvErr "%10.*f\n", Precision+1, SumM / N);
fprintf(Of, F_RelErr " %4.2f\n", (SumD ? SumM / SumD : 0.0 ));
fprintf(Of, F_CorrCoeff " %4.2f\n",
( SumRP -SumR * SumP / N ) /
( sqrt( (SumRR - SumR * SumR / N) *
(SumPP - SumP * SumP / N) ) + 1E-6 ));
/* Free local storage */
ForEach(i, 0, MaxCase)
{
Case[i] = Blocked[i];
}
Free(ErrMag);
Free(Blocked); Blocked = Nil;
}
/*************************************************************************/
/* */
/* Prepare data for cross-validation */
/* */
/*************************************************************************/
void Prepare(void)
/* ------- */
{
CaseNo i, First=0, Last, *Temp, Hold, Next=0;
int Bin, Group;
ContValue MinVal, MaxVal, Range;
Temp = Alloc(MaxCase+1, CaseNo);
ForEach(i, 0, MaxCase)
{
Temp[i] = i;
}
Shuffle(Temp);
/* Find maximum and minimum class value */
MaxVal = MinVal = Class(Case[0]);
ForEach(i, 1, MaxCase)
{
if ( Class(Case[i]) < MinVal ) MinVal = Class(Case[i]);
else
if ( Class(Case[i]) > MaxVal ) MaxVal = Class(Case[i]);
}
Range = MaxVal - MinVal;
/* Sort into groups of approximately the same value */
ForEach(Bin, 0, 9)
{
Last = First - 1;
ForEach(i, First, MaxCase)
{
Group = Min(9, 10 * (Class(Case[Temp[i]]) - MinVal) / Range);
if ( Group == Bin )
{
Last++;
Hold = Temp[Last];
Temp[Last] = Temp[i];
Temp[i] = Hold;
}
}
First = Last+1;
}
/* Organize into stratified blocks */
ForEach(First, 0, FOLDS-1)
{
for ( i = First ; i <= MaxCase ; i += FOLDS )
{
Blocked[Next++] = Case[Temp[i]];
}
}
Free(Temp);
}
/*************************************************************************/
/* */
/* Shuffle the data items */
/* */
/*************************************************************************/
void Shuffle(int *Vec)
/* ------- */
{
int This=0, Alt, Left=MaxCase+1, Hold;
ResetKR(KRInit);
while ( Left )
{
Alt = This + (Left--) * KRandom();
Hold = Vec[This];
Vec[This++] = Vec[Alt];
Vec[Alt] = Hold;
}
}
xval 0000755 0000764 0000764 00000005724 11340612250 011670 0 ustar quinlan quinlan #! /bin/csh
#---------------------------------------------------------------------
# Multi F-fold cross-validation script
#---------------------------------------------------------------------
#
# Invocation:
# xval [Cubist options] [F=folds] [R=repeats] [+label] [+d]
#
# Carries out R F-fold cross-validations
#
# If +d is used, individual results from each block are left in
# .o[+label]
# Summary of cross-validations is written to
# .res[+label]
#---------------------------------------------------------------------
# Sort the options into those applying to Cubist and the rest
set opts =
set folds = 10
set repeats = 1
set label =
set filestem = undefined
set i = 1
while ( $i <= $#argv )
set opt = $argv[$i]
switch ( $opt )
case "F=*":
set folds = `echo $opt | sed s/F=//`
breaksw
case "R=*":
set repeats = `echo $opt | sed s/R=//`
breaksw
case "+d":
set details
breaksw
case "+*":
set label = $opt
breaksw
case "-f":
@ i++
set filestem = $argv[$i]
breaksw
case "-f*":
set filestem = `echo $opt | sed s/-f//`
breaksw
case "-r":
case "-n":
case "-e":
case "-S":
case "-I":
case "-C":
@ i++
set opts = ( $opts ${opt}$argv[$i] )
breaksw
case "-r*":
case "-n*":
case "-e*":
case "-S*":
case "-I*":
case "-C*":
set opts = ( $opts $opt )
breaksw
case "-i":
case "-a":
case "-u":
set opts = ( $opts $opt )
breaksw
case "-X":
@ i++
set folds = $argv[$i]
breaksw
case "-X*":
set folds = `echo $opt | sed s/-X//`
breaksw
default:
echo "Unrecognised or inappropriate option" $opt
case "-h":
echo ""
echo "Summary of options for xval:"
echo ""
echo " F= set f folds"
echo " R= repeat r times"
echo " +d retain detailed files"
echo " +s label all output files with suffix +s"
echo " -f set application name"
echo " -r set maximum number of rules"
echo " -e set maximum extrapolation"
echo " -i use instances and rules"
echo " -a allow use of instances with rules"
echo " -n set number of nearest neighbors (1-9)"
echo " -S set training sample percentage"
echo " -I set random seed [ignored]"
echo " -X cross-validate"
echo " -C committee model"
echo " -h print this message"
exit 1
endsw
@ i++
end
# Clear the predictions file
cp /dev/null $filestem-tmp.pred
# Repeat cross-validations, incrementing the random seed
set r = 0
while ( $r < $repeats )
set outf = $filestem.o$r$label
cubist -f $filestem $opts -X $folds -I $r >$outf
cat $filestem.pred >> $filestem-tmp.pred
@ r++
end
# Remove the temporary files and summarize results
mv $filestem-tmp.pred $filestem.pred
summary -f $filestem -n >$filestem.res$label
if ( ! $?details ) rm -f $filestem.o[0-9]*$label