Author: Graham Laight
Date: 13:23:20 06/29/00
Go up one level in this thread
On June 28, 2000 at 16:16:30, Bruce Moreland wrote: >On June 28, 2000 at 13:58:28, Graham Laight wrote: > >>I assume that what you're talking about is a database of piece patterns that can >>occur in a chess game, together with some indication of what these patterns >>mean. >> >>Unfortunately, when we try to suggest that a NN is the best AI instrument for >>abstract pattern recognition, all we seem to hear is that NNs are quite >>incapable of being trained to recognise chess patterns - as if the mechanism of >>the human brain is some sort of magic... > >It's easy to write an algorithm that can tell the difference between an X and an >O. > >It's harder to write one that tell the difference between the a bunch of good >guitar players, and tell you why one of them is the best, and why one solo was >his best solo. The whole idea of using tools like NNs, genetic algorithms and fuzzy logic is that they CAN be used to attack problems that are difficult to write algorithms for. If a need to rate guitar music arose, which tool would you reach for - a NN package, or a C++ compiler? >I suspect that chess is more like the latter problem than the former. Exactly. You seem to be debating FOR the proposal! >Humans see patterns in positions, but they are also very good at weighting the >values of these patterns in specific cases, and they back this impressional >knowledge up with good calculating ability. Yes - very true! And if we want the computer to do the same, we're going to have to give it some of what a human has. >I encourage other approaches, of course. If they can be made to work, it would >be significant in many ways. It would be nice if it could actually happen >before the centennary of the work of Turing and Shannon though. > >Personally, if I had to pick a multi-generational problem to work on, it would >be colonization of a nearby star system, rather than a "pure" AI approach to >computer chess. I suspect that if the NN approach to doing chess (in my view, an NN to recognise the existence of each type of pattern in the presented position, and one big (or a lot of small) NNs to make an overall evaluation based on the outputs of the individual NNs) had received as much development effort as the prevailing doctrine (write evaluation functions directly in a language like C), then by now we'd be able to evaluate positions more accurately (if not more quickly) than we actually can. -g >bruce
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.