Not logged inRybka Chess Community Forum
Forum RybkaChess Help Search Register Login
Up Topic Rybka Support & Discussion / Rybka Discussion / Directed Evolution for Evaluation tuning
- - By XmikeX (**) [us] Date 2007-09-13 05:22
From authors, we sometimes see the following :

"Yeah, that's deficiency X.  If I fix X, then I end up hurting Y and Z, and I don't know how to fix X otherwise."

What are some thoughts on genetic algorithms to tune engine evaluations, especially in a case like the above?

Set up success/fail parameters, give it an algo, crunch for a few billion x billion x billion iterations, open the lid, see if anything crawls out...

That sort of thing..

?

*- Just found this student article -> http://www.cs.utah.edu/~wyman/classes/ML_proj/paper.html   "Using Genetic Algorithms to Learn Weights for Simple King-Pawn Chess Endgames"
Parent - By lkaufman (*****) Date 2007-09-13 13:40
The use of genetic algorithms is already proving to be the best method in the game of Go. In my opinion, chess programs will adopt it once the number of processors reaches the point of significant diminishing returns, which will probably be fairly soon.
Parent - By Fulcrum2000 (****) [nl] Date 2007-09-13 16:54
I read the paper and I think it one of the most cumber sum methods of 'creating chess knowledge'. Also the results obtained (~25% score in a dozen or so well trained boards is not very encouraging). I think that genetic algorithms for chess will take many more decades to be useful to even a tiny degree. Chess is simply to complex with the different behavior of the pieces (pawn/rook/knignt etc), and even the same piece in a different position (rook before/after rocade).
Parent - By lkaufman (*****) Date 2007-09-14 03:35
Sorry, I must have been groggy, my response about GO referred to Monte Carlo methods, not genetic algorithms. I should have replied that genetic algorithms are probably promising for tuning portions of an eval function, but not the whole eval function at once as there are too many terms. My former partner in "Rexchess" and "Socrates", Don Dailey, is researching both genetic algorithms and Monte Carlo in chess and GO respectively.
Up Topic Rybka Support & Discussion / Rybka Discussion / Directed Evolution for Evaluation tuning

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill