Author: Ratko V Tomic
Date: 05:19:39 10/25/99
Go up one level in this thread
> I think this is a very insightful post. Well, thanks, although I think the insightful part is probably due to the Botvinik's ideas (which I expressed above) on how the future higher species of the chess programs will work i.e. via the multilayered trees and their control modules, with chess knowledge residing mainly in the higher layers (and not in the low level alpha-beta searchers as it is done today). Anyone who has used programs to analyse in depth a position knows that letting the program work completely on its own for a day or two isn't the best way. Instead we all try out various ideas by guiding the program and letting it clear up the tactical aspect of the variations. In much shorter time we can get a very good picture what is possible in a given position and how would it work out on longer run, more accurate than a huma alone or program alone would produce. Therefore, we know that, at least in principle, a better way exists to use alpha-beta. The higher level guidance provides gain not only through its use of a sparse higher level tree (with regular alpha-beta working in between the nodes of that tree) of practically unrestricted depth, but it also makes the low level searcher much more efficient since it adds high level constraints on the search (this is the same kind of efficiency gain that a pure mate searcher has over the general purpose playing program searcher). The low level alpha-beta search can be initiated multiple times from the same high level node, but with different (strategic/tactical) objectives each time. This is much more discriminating (and harder to fool with delaying moves) way to analyse than merely adding in an incoherent amorphous sum all the positional and material contributions at each terminal node.
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.