Author: Dann Corbit
Date: 14:53:04 04/09/04
Go up one level in this thread
On April 09, 2004 at 16:49:43, Uri Blass wrote: >On April 09, 2004 at 16:24:42, Dann Corbit wrote: > >>I have been thinking about forward pruning. (Did you smell wood burning?) >> >>The ten cent description of Baysian logic to those who have not examined it: >>As more information comes in, we revise our probability estimates. The Monty >>Hall problem is an excellent example of it. >> >>Anyway, when you look at the techniques used to decide whether or not to >>exercise some sort of forward pruning that are not complete no-brainers like >>Alpha-Beta cutoffs, it seems logical to me to employ Baysian logic. The reason >>is that advancing search depths give increased information. >> >>It seems a perfect fit for the theory. >> >>It seems to me it could even be used with a notion like: >>Given the large number of available moves and the huge negative score, do we >>even need to verify this null move? >> >>And things of that nature. >> >>Has anyone tried it? > >I do not understand what you suggest. This is what made me think about it: http://www.enm.bris.ac.uk/teaching/enjfb/emat31600/Fuzzy%20Decision%20Trees.ppt >If you talk about using probabilities for pruning decisions and to prune moves >with big probability to fail low then I agree that it may be a good idea but the >problem is how to evaluate probabilities. > >I use some statistics about history fail high history fail low in movei forward >pruning and I have no doubt that it can be improved. Suppose (for instance) that we have some branch that loses a whole piece. Perhaps there is a way to formulate a rule about how deeply to explore that branch. If we set some hard limit (like the program must always search at least 1/2 of the root level search for every branch) then we can still produce an algorithm that always produces correct results (albeit perhaps much more slowly than the less speculative algorithm).
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.