Author: Peter Fendrich
Date: 14:18:02 06/29/04
Go up one level in this thread
On June 29, 2004 at 13:07:24, Dan Andersson wrote: > What test methodology do you use? And what kind of GA? And have you considered >GP and GEP? Just curious :) > >MvH Dan Andersson Mvh - Are you sure everyone knows the meaning of this? :-) I'm doing this in several steps. Right now: GA for evaluation weights just to warm up a bit. I don't exactly expect it to give me much except maybe some improved values. Each individual is one instance of "all" evaluation weights. I load it into the program and run. I have a file from Dann with 89000 quiet epd-positions where several top engines agree on the bestmove. The reward is based on bestmove hits. Type: I don't know if "4 monkey's in a bus" tell you anything... I have classical mutation and 1-point crossover plus Gaussian mutation. After summer: GP for search parameters and here is where the challenge really is... My setup, very short (for GP): I've got a file with about 14000 high quality games from Dann. Step 1, creating learning data. I want to go through all games with some (4?) high rated engines and produce an epd-file with about the same content as the epd-file in the GA test. A move is higher rated if more engines agree on it. I think the c0 tag can be used. Like this: r2qkb1r/pp1n1ppp/2p2n2/3pp2b/4P3/3P1NPP/PPPN1PB1/R1BQ1RK1 b kq e3 c0 Fritz: c4, CM9K: c4, ChessTiger: g4; Step 2, use the epd-file. The set of GP operations will include some of my own search functions and parameters like the R-value for Nullmoves, extension logic etc. The reward will be based on getting result in as few nodes as possible. For step 2 I will need soooo much cpu and hope for some good help and a distributed method to collect the results that I haven't thought of yet.... /Peter
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.