gpdata_gp96.readme

标准的GP源代码,由Andy Singleton维护

W. B. Langdon							21 March 1995

TEST CASES FOR GP-96 PAPER "USING DATA STRUCTURES WITHIN GENETIC PROGRAMMING"

This file documents three data files each of which contains the test
case for one of the experiments described in the paper "Using Data
Structures within Genetic Programming" by W. B. Langdon to be
presented at GP-96. These test cases are made available so as to make
it easier for you to reproduce the experiments described. Please feel
free to contact me if you have any questions.

A postscript version of the paper is available via anonymous ftp from
node cs.ucl.ac.uk in directory genetic/papers file WBL.gp96.ps The
paper describes GP solving three ``classic'' problems. The first two
require the GP to distinguish between sentences that are in a language
and those that are not given positive and negative training examples
of the language. 

The two languages are, correctly nested brackets and a Dyck language
(correctly nested brackets of different types).

The third problem is to evaluate integer Reverse Polish (postfix)
expressions.

	W. B. Langdon,				Phone [phone]
	Computer Science Dept.			Fax   [phone]
	University College London,
	Gower Street,				[e-mail]
	London, WC1E 6BT
	England         

	http://www.cs.ucl.ac.uk/staff/W.Langdon/
----------------------------------------------------------------------

Files 
r1.7  gp.test.brack  (Nested brackets)
r1.15 gp.test.karp   (Dyck problem)
r1.12 gp.test.rpn    (Reverse Polish Expression)

Format of data 
% indicates start of comment which finished at the end of the line

GPtests nnn	nnn indicates max size of test case

stop_on nnn	nnn is a bit mask. Each bit indicates if the testing
		may continue following a program indicating a particular
		exception. These are explained in the paper and
		follwing comment 

start nnn	Start of a sequence of tests. nnn is a seed used to
		initialise indexed memory. 0 indicates initialise to zero.

r0..r9		writable registers

tree rN xx	Call tree (ie run the GP program under test using from
		tree). Typically tree will be "rpb", ie run the main
		result producing branch.  rN indicates a register
		in which to store the value returned by the program and
		xx is the input to the program. This may be an integer
		or the contents of a register.

		If the tree takes no arguments xx is omitted, if it
		has two arguments then xx would be followed by another
		value or register. 

		If the tree doesn't return a value rN is omitted

end_test	end of test sequence

FINISHED_TESTS	Whole test case is finished

A number of statements allow specific conditions to be tested
for. These are logically combined by the next break statement. Ie a
test is only passed if ALL tests since the last "start" or "break" have
been passed.  If any test fails then the "stop_on_error" on exception
bit (1024) is set. If this bit is set in the stop_on mask the testing
of this subsequence will stop at the next break

break		Iff previous tests are true then 
			increment the number of passes
		else
			may stop testing (depending on stop_on_error)

		Testing will stop if following an exception (eg
		loop_too_nested_error) if the corresponding bit was
		set by stop_on

scorneq_ rN ss	if(NOT rN) increase score by ss. In these tests ss is
		always one. 

		In the first two experiments the programs
		score is the number of "scorexx"/"break" it passes.

scoreq_ rN ss	if(rN) increase score by ss (always one). 

scoreq rN xx	if(rN==xx).

		In the reverse polish experiment a separate score is
		kept for each of the five main trees. Each time a
		"scorexx"/"break" is passed the score for each tree
		executed since the previous break (or start) is
		increased by one. 

hurdle pp	Do not run any more test sequences, unless pp or more
		breaks have been passed.