Revision 305d3ef4ad2d0305ba48507be544b10256b0b90b authored by Lee Spector on 02 November 2013, 15:28:39 UTC, committed by Lee Spector on 02 November 2013, 15:28:39 UTC
1 parent bf95b4b
old-version-history.txt
VERSION HISTORY -- NOW OBSOLETE. See https://github.com/lspector/Clojush/commits/master
20100227: - First distributed version.
20100301: - Added (shutdown-agents) for proper termination.
20100306: - Added history (of total errors of ancestors) to individuals.
- Commented out (shutdown-agents) because it prevents multiple
runs in a single launch of a Clojure REPL.
20100307: - Fixed bug in history: reproductive auto-simplification added
was adding duplicate items.
20100314: - Added instructions: *_shove, code_extract, code_insert,
code_subst, code_contains, code_container, code_position,
code_discrepancy, *_rand. NOTE that the presence of *_rand
instructions means that programs produced using the full set
of instructions may be non-deterministic. As of this writing
pushgp (via evaluate-individual) will evaluate an individual
only once, so it will always have whatever fitness value it
had upon first testing.
- Added globals to support integer_rand and float_rand:
min-random-integer, max-random-integer, min-random-float
max-random-float.
- Fixed bug in code_car that could produce nil.
- Made execute-instruction safe for nil (although it shouldn't
arise anyway).
- Added stress-test for testing and debugging new Push
instructions. See the stress-test documentation string for
details.
- Implemented size limits on intermediate results on code stack
(in code_cons, code_list, code-append, code_insert, code_subst,
exec_s, exec_y).
- Fixed bug in exec_s (was always a noop previously).
20100319: - Added problem-specific-report function (to be customized per
problem). This can also be a convenient place to put other
stuff that you want done once per generation.
- Added support for saving lists of ancestors (maternal line
only) along with global parameters to turn both this and the
saving of total-error histories on and off.
- Added missing calls to "keep-number-reasonable" in numeric
Push instructions. This eliminates some potential crashes from
runaway number growth.
20100320: - Added print-ancestors-of-solution parameter and code.
- Print simplification in report only with non-zero value for
report-simplifications parameter.
- Added sextic polynomial regression example. This example also
demonstrates the use of fitness penalties for abnormally
terminating programs.
- Added a new argument to problem-specific-report (NOTE: not
backward compatible).
20100417: - Added thread-local random number generator objects to avoid
contention.
- Print parameters at the start of pushgp.
- Added readme comments about concurrency, -Xmx2000m, and
-XX:+UseParallelGC.
- Converted time limit code to use System/nanoTime (thanks to
Brian Martin). This means that time limits must now be
expressed in nanoseconds rather than milliseconds, and I
believe it will eliminate contention for shared Date objects
(but this should be checked; if there is contention then
we should revert to using Date and use thread-local date
objects as is being done with the random number generator
objects).
- Added print-return utility function for debugging.
- Added a new Push instruction, code_wrap, which pushes a 1-item
list containing the previous top item of the code stack.
- Added a new Push instruction, code_map, which acts much like
Lisp's (or Scheme's or Clojure's) "map" functions, using the
top item on the exec stack as the function to map and the top
item on the code stack as the list to map it down. The list of
results is left on top of the code stack. This is implemented
as a "macro" instruction that expands into a Push code
fragment that: 1) for each item in the list on top of the
code stack (or for the single non-list item that is there)
quotes the item onto the code stack and then runs the code
that's on top of the exec stack; 2) uses code_wrap to push a
list containing the last result onto the code stack; 3)
executes as many instances of code_cons as are necessary to
add all of the other results onto the list. Note that this
will act like an ordinary "map" function only when the code on
the exec stack leaves a single output on the code stack in
place of each input on the code stack; if it consumes or
produces more or less code then the effect will be quite
different.
20100502: - Made thread-local random integer function (lrand-int) safe for
bignums, but arguments greater than 2^31-1 are treated as if
they were 2^31-1 (2147483647).
20100526: - Reimplemented subst to use clojure.walk/postwalk-replace. Also
fixed comment, which described args backwards. (The behavior
was correct, emulating Common Lisp subst.)
20100918: - Created Eclipse project.
- Deleted re-load/backtrace utils.
- Removed shuffle, as it is now in clojure.core (in Clojure 1.2).
- Removed gratuitous def in define-registered.
- Added atom for instruction-table.
- Added atom for registered-instructions; NOTE: requires user
code that refers to registered-instructions to refer to
@registered-instructions instead. (Example odd.clj changed
to reflect this.)
- Added to-do item "Convert structs to records, which should be
faster. Experiments with Clojure 1.2 show this to be faster
but there are not good examples yet to serve as the basis for
changes.
- Added atoms for global-atom-generators and
global-max-points-in-program.
- Changed pushgp to take keyword arguments rather than a parameter
map. NOTE: this requires calls to pushgp to be written differently.
Updated examples to reflect this.
20100921: - Removed random-element in favor of rand-nth.
- Cleaned up indentation, miscellaneous other cosmetic things.
- Added namespaces for all example files.
- Updated README to mention requirement for clojure 1.2 and to
remove mention of ClojureX which has been discontinued.
- Converted structures to records (a clojure 1.2 feature, should
be faster).
20101005: - Added error-handlers to agents.
20101014: - [Artificial ant, krypto, tg8, decimation]
- Added articial ant example from Koza (via Luke).
- Added "tg8" integer symbolic regression problem.
- Added krypto number puzzle example.
- Added pushgp "decimation" feature, in which elimination
tournaments, conducted after fitness testing, reduce the
size of the population to a specified fraction of its original
size (specified in a decimation-ratio argument to pushgp;
the tournament sized is specified via decimation-tournament-size).
The ordinary tournament-size parameter is still used for subsequent
selection from the decimated population. Any specified trivial
geography applies both to decimation and to subsequent selection.
20101017: - Reverted from records to structs; wasn't significantly faster and
structs allow for greater flexibility in use of state as map.
20101102: - Switched to new clojure.core/frequencies from depricated
seq-utils/frequencies (h/t Kyle Harrington), and similarly
for flatten.
- Added :include-randoms keyword argument for registered-for-type.
Defaults to true, but if specified as false then instructions
ending with "_rand" will be excluded.
- Raised invalid output penalty in tg8 (it was lower than reasonable
errors for that problem).
20101103: - Converted evalpush-limit and evalpush-time-limit into vars
(global-evalupush-limit and global-evalpush-time-limit) bound
to atoms that are reset by calls to pushgp (keyword arguments
:evalpush-limit and :evalpush-time-limit).
- Changed pushgp parameters in the tg8.clj example.
20101104: - Implemented stackless tagging (see http://push.i3ci.hampshire.edu/
2010/10/28/stackless-tagging/). Tag instructions take one of
the following forms:
tag_<type>_<number>
create tage/value association, with the value taken from the
stack of the given type and the number serving as the tag
untag_<number>
remove the association for the closest-matching tag
tagged_<number>
push the value associated with the closest-matching tag onto
the exec stack (or no-op if no associations).
Here "closest-matching" means the tag with lowest number that
is greater than or equal to the number of the tag being matched,
or the lowest-numbered tag if none meet this criterion. Tag
instructions are not implemented in the same way as other
instructions; they are detected and handled specially by the
interpreter (see execute-instruction). Tag instructions
can be included in pushgp runs by using the new ephemeral
random constant functions tag-instruction-erc,
untag-instruction-erc, and tagged-instruction-erc, each of
which takes a limit (for the numerical part) as an
argument.
- Added examples using tags: tagged_ant, tagged_regression, and
tagged_tg8.
20101106: - Tweaked parameters in ant examples; among other things,
increased simplification since bloat was an issue. Also
added some evolved solutions in comments.
20101107: - Added Koza's lawnmower problem example; this demonstrates how
to add a new type/stack on a problem-specific basis, without
altering clojush.clj.
20101204: - Added pushgp-map, which allows pushgp calls on maps of arguments,
and a demonstration of its use in argmap_regression.clj.
- Added :gen-class and -main definition (thanks Kyle Harrington).
- Fixed eval-push termination to return :abnormal for exceeding
time-limit (thanks Kyle Harrington).
20101205: - Added modified version of Kyle's version of the intertwined
spirals problem.
- Minor changes to this README.
20101208: - Added alternative methods for node selection, used in mutation
and crossover (drafted by Brian Martin, with suggestions
from Kyle Harrington). This introduced three new globals:
global-node-selection-method, global-node-selection-leaf-probability,
and global-node-selection-tournament-size, each of which holds
an atom, and three new parameters to pushgp: node-selection-method,
node-selection-leaf-probability, and node-selection-tournament-size.
The node-selection-method can be :unbiased (in which case nodes
are selected using the uniform distribution that was previously
used -- this is the default), :leaf-probability (in which case
the value of the node-selection-leaf-probability argument,
which defaults to 0.1, specifies the probability that leaves,
as opposed to internal nodes, will be selected -- this is the
method used by Koza and others in tree-based GP), or
:size-tournament (in which case the value of the
node-selection-tournament-size argument, which defaults to 2,
determines the tournament size for node tournaments, with the
largest subtree in the tournament set being selected).
20110111: - Added zipper stack and functions (thanks to Kyle Harrington for
draft code, although this was re-written).
- Added registered-nonrandom function.
- Modified odd.clj example to use registered-nonrandom.
- Added examples/dsoar.clj, a version of the "Dirt-Sensing,
Obstacle-Avoiding Robot" (DSOAR) problem first described in:
Spector, L. 1996. Simultaneous Evolution of Programs and their
Control Structures. In Advances in Genetic Programming 2, edited
by P. Angeline and K. Kinnear, pp. 137-154. Cambridge, MA: MIT Press.
http://helios.hampshire.edu/lspector/pubs/AiGP2-post-final-e.pdf
This version was written by Brian Martin in 2010-2011.
20110118: - Added support for tagged_code_<number> instructions. These are like
tagged_<number> instructions except that retrieved values are pushed
onto the code stack rather than the exec stack. Without these the
only way to get tagged values onto the code stack is to wrap
values with code_quote prior to tagging. An alternative approach
is to add new tagging instructions that automatically perform
code_quote wrapping, but for full generality that would require
new instructions for each type; also quote-tagged values would
always be destined for the code stack, while the scheme adopted
here allows any stored value to be retrieved either to exec or to
code.
- A value of 0 for the evalpush-time-limit parameter of pushgp
now means that no time limit will be enforced. This is also
now the default.
20110322: - Tag reference bug fixed in closest-association (>= changed to <=).
- Added mux (multiplexer) example (a couple of approaches in one file).
20110409: - Added support for no-pop tagging through a var called
global-pop-when-tagging (holding an atom with a boolean value)
and a boolean argument to pushgp called pop-when-tagging.
The defaults are true, for backwards compatibility. If
@global-pop-when-tagging is false (which will result from
passing false as a :pop-when-tagging keyword argument to pushgp)
then executing instructions of the form tag_<type>_<number>
will tag a value as usual, but the tagged item will not be popped
from its source stack.
- Removed no-pop hackage from mux example (thanks Kyle).
20110424: - Added Gaussian mutation for floating point literals. This is
a genetic operator on par with ordinary mutation, crossover,
and simplification, with the probability for applying this operator
set with the gaussian-mutation-probability argument to pushgp
(which defaults to zero). The behavior of this operator, when used,
is controlled by two other arguments to pushgp,
gaussian-mutation-per-number-mutation-probability (which is the
probability that any particular floating point number in the
program will actually be mutated -- this defaults to 0.5) and
gaussian-mutation-standard-deviation (which specifies the standard
deviation of the Gaussian noise generator that is used to
produce changes to numbers -- this defaults to 0.1).
- Added examples/gaussian_mutation_demo.clj to demonstrate Gaussian
mutation.
- Added examples/korns_regression_p12.clj, a symbolic regression
problem based on Michael Korns's draft chapter from GPTP11.
20110505: - Added complex number support. New instructions for the 'complex'
stack include: pop, dup, swap, rot, flush, eq, stackdepth, yank,
yankdup, shove, rand, add, sub, mult, divide, fromfloat,
frominteger, fromfloats, fromintegers, conjugate, magnitude,
and principal_sqrt. (Brian Martin)
20110517: - Added a "scaled-errors" function to support error-scaling as
described by Maarten Keijzer in Scaled Symbolic Regression, in
Genetic Programming and Evolvable Machines 5(3), pp. 259-269,
September 2004. This must be used in a problem's error function,
and then the outputs of the evolved program must be "unscaled."
See the documentation string for scaled-errors and also
examples/scaled_sextic.clj for details.
- Added examples/scaled_sextic.clj to demonstrate the use of
scaled-errors.
- Changed examples/sextic.clj to use squared errors and an error
threshold, in order to facilitate comparisons between the
versions that do and don't use error scaling.
- Made minor changes to the korns_regression_p12 example.
20110526: - Enforce size limits on zipper manipulation results.
20110607: - Added overlap utility function, overlap-demo (which just prints
some examples to show how overlap works), and code_overlap
instruction. Overlap can be applied to any two (possibly
nested) things and it returns a number between 0 (meaning
no overlap) and 1 (meaning exact match). The overlap utility
function returns a ratio, but the code_overlap instruction
pushes a float.
- Removed complex number support from 20110505. There were previous
reports of problems and I've noticed problems from the fact that
(apply + ()) => zero (as opposed to 0) in the clojush namespace
defined by the code as revised for complex number support. If
someone knows how to re-introduce complex number support without
such problems then please let me know.
20110618: - Switched to Kyle Harrington's version of overlap; it's more clear,
possibly faster, and may fix a hard-to-trace bug that cropped up
in a long evolutionary run (?).
20110624: - Replaced lawnmower and dsoar examples with bugfixed versions
(thanks to Kyle Harrington).
- Added namespace and made miscellaneous other changes to
clojush-tests.clj.
- Added support for tagged-code macros. Tagged-code macro calls
have the effect of code instructions, but they take their
code arguments from the tag space and they tag their code return
values. They are implemented as macros to leverage the existing
code instruction set; note that this means that a single call
will contribute more than one iteration step toward the
evalpush-limit. Tagged-code macros appear in programs as hash maps
that map :tagged_code_macro to true, :instruction to a code
instruction, :argument_tags to a sequence of tags to be used
to acquire arguments, and :result_tags to a sequence of tags
to be used for tagging results. Execution of a macro expands
the needed code onto the exec stack to grab arguments from the tag
space and push them on the code stack, execute the code instruction,
and tag results. Note that results will also be left on the code
stack if global-pop-when-tagging is false. Conceptually, tag values
are "baked in" to these macros in much the same way that tag values
are "baked in" to the instruction names for stackless tag
instructions; we use hash maps simply because there is more
information to bake in and this prevents us from having to parse
the names (which would get messy and also waste time). Because
the maps make it difficult to read programs, however, a utility
function called abbreviate-tagged-code-macros is provided to
produce copies of programs with more human-readable (but not
currently executable) representations of tagged-macro calls.
A tagged-code-macro-erc is provided to generate random tagged-code
macros in pushgp runs. A new example, codesize20, provides
a simple demonstration of the use of tagged-code macros.
- Replaced walk-based code-manipulation with walklist functions
that only traverse list structure. This fixes an interaction
between map literals (e.g. tagged-code macros) and program
structure.
20110629: - Fixed abbreviate-tagged-code-macros printing of empty lists.
- Added seq condition to walklist to permit walking of seqs that
aren't actually full-fledged lists.
20110702: - Several fixes/refinements to tagged-code macros:
- Fixed incorrect no-op of arg-free calls with empty tag space.
- Added :additional_args to tagged-code macro structure; the
value should be a list of items and these will be executed
in order before calling the macro's instruction.
- Added optional 5th arg to tagged-code-macro-erc; this should
be a function of zero args that will be called to produce
the value of :additional_args (e.g. if you want to have one
random integer arg then you could specify a 5th arg of
(fn [] (list (lrand-int 101))).
- Changed format produced by abbreviate-tagged-code-macros to
handle :additional_args and to be slightly more concise.
20110714: - Added "trace" argument to eval-push and run-push. If this is
true then the resulting state will map :trace to a list of
executed instructions and literals, in reverse order of
execution. If the argument is :changes then instructions that
have no effect on the state will be excluded.
20110809: - Several additions/enhancements by Kyle Harrington:
- Converted problem-specific-report to a parameter in pushgp.
- Added reporting of program repeat counts in population.
- Added "error-reuse" parameter to pushgp for use in stochastic
and dynamic problems (for which reuse would be turned off).
- Added examples/mackey_glass_int.clj, a symbolic regression
problem as described in Langdon & Banzhaf's 2005 paper
(citation in file).
- Added examples/pagie_hogeweg.clj problem, a difficult
symbolic regression problem when coevolution is not used.
Introduced by Pagie & Hogeweg's 1997 paper (citation in file).
20110911: - Switched from Eclipse to Clooj/Leiningen for development and
rearranged the project for this.
- Added calls to end of all example files.
- Examples can be run from the OS command line (assuming
that leiningen is available) with calls like:
lein run examples.simple-regression
- Added local-file dependency and used file* for file access.
- Removed ant and tagged-ant examples because of bugs related
to confusion of push interpreted states and ant world states.
20111104: - Added string stack and a variety of string stack instructions.
- Added two example pushgp runs that use the string stack in
the file examples/string.clj.
20111112: - Obsoleted this version history in favor of the commit logs
at https://github.com/lspector/Clojush/commits/master
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...