Author: Danniel Corbit
Date: 18:07:39 05/19/98
Go up one level in this thread
On May 19, 1998 at 10:36:33, Don Dailey wrote: >I just recently posted about this and briefly described some >experiments I did and what I thought the problems were. I am >generally "cautiously optimistic" about this approach. There >are some problems that need to be solved, particulary with >the "fitness" function to make this work well. I commented >that it's a general problem in computer chess, how to measure >small improvements to your chess program. > >John Stanback played around with this too. Both of us got >very reasonable numbers for piece values. I can give you >more details and observations if you are interested. Also >I'm interested in hearing your ideas. I have a notion to create a chess engine with a threaded virtual analysis function. We can have one thread running alpha-beta requests. One thread running bayesian requests. One thread running neural net forward propagation requests. One thread running neural net backward propagation requests. Now, most of the requests will overlap, so the added effort will not be great if we can share the individual move values in a common table. The different numbers created by these independent evaluators can be weighted. An additional function can accept all of these inputs and decide on a move. It seems to me that they all have different strengths. Bayesian and Alpha-Beta are tactical, but in different ways. Neural approaches are strategic, as they recognize favorable patterns, like bad inactive bishops or strong pawn chains. The combined information should produce a very strong evaluation, if combined correctly. Again, I have never written a chess program. These are just notions that seem sensible to me from what I have read so far. What do 'real' chess programmers think of this notion?
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.