syllogismos · August 29, 2015 14:00
diff --git a/genetic-programming-brief-explanation.txt b/genetic-programming-brief-explanation.txt
 http://forums.aichallenge.org/viewtopic.php?f=17&t=1136

 We represent every program internally as an object tree,
 very much like a compiler internally represents a program
 after parsing and before generating executable code. 

 A compiler manipulates these objects in order to optimize
 the code, we do it in order to randomly change the
 code (mutation) and to insert parts of one program
 into another (crossover).
 For example, an IF node has 3 subnodes: a condition,
 a then part and an else part. When we do a crossover,
 we randomly chose any node (say the condition) and replace
 it by any randomly selected boolean expression that appears
 anywhere in the code of the other program. The result is a
 new program that does nonsense in most cases, but occasionally
 it works better than both of the original programs.
 Doing these object manipulations is actually pretty simple,
 just copying references to objects while making sure the 
 result is a valid program.

 In an effort to make this process more efficient, we analyse
 the code before we use it and remove all the useless parts
 whenever we can identify them. For example, if we see

 Code:  IF 4 > 7 THEN <a> ELSE <b>
 we replace it by
 Code: <b>

 If we see 
 Code: IF planet_is_mine AND planet_is_mine THEN ...
 we replace it by
 Code: IF planet_is_mine THEN ...

 If we see
 Code: IF planet_is_mine AND planet_is_neutral THEN ...

 then we don't recognize in our static analysis that this
 is nonsense. However, we insert code that counts how often
 this combined condition is true. If it appears to be never
 true, then we remove the whole statement. If it appears to
 be always true, then we stop checking it.

 The main purpose of these simplifications is that we want
 to copy some code from a parent to a child only when it actually
 contributed to the success of the program. Propagating dead
 code is much less likely to improve the strategies than
 propagating code that has been used to win games.
	http://forums.aichallenge.org/viewtopic.php?f=17&t=1136

	We represent every program internally as an object tree,
	very much like a compiler internally represents a program
	after parsing and before generating executable code.

	A compiler manipulates these objects in order to optimize
	the code, we do it in order to randomly change the
	code (mutation) and to insert parts of one program
	into another (crossover).
	For example, an IF node has 3 subnodes: a condition,
	a then part and an else part. When we do a crossover,
	we randomly chose any node (say the condition) and replace
	it by any randomly selected boolean expression that appears
	anywhere in the code of the other program. The result is a
	new program that does nonsense in most cases, but occasionally
	it works better than both of the original programs.
	Doing these object manipulations is actually pretty simple,
	just copying references to objects while making sure the
	result is a valid program.

	In an effort to make this process more efficient, we analyse
	the code before we use it and remove all the useless parts
	whenever we can identify them. For example, if we see

	Code: IF 4 > 7 THEN <a> ELSE <b>
	we replace it by
	Code: <b>

	If we see
	Code: IF planet_is_mine AND planet_is_mine THEN ...
	we replace it by
	Code: IF planet_is_mine THEN ...

	If we see
	Code: IF planet_is_mine AND planet_is_neutral THEN ...

	then we don't recognize in our static analysis that this
	is nonsense. However, we insert code that counts how often
	this combined condition is true. If it appears to be never
	true, then we remove the whole statement. If it appears to
	be always true, then we stop checking it.

	The main purpose of these simplifications is that we want
	to copy some code from a parent to a child only when it actually
	contributed to the success of the program. Propagating dead
	code is much less likely to improve the strategies than
	propagating code that has been used to win games.
No results found