Author: guy haworth
Date: 07:50:13 06/23/00
Go up one level in this thread
I suggested the chess endgame to a UK student as a suitable domain to demonstrate ANNs 'training on' and successfully evaluating a set of data. With Ken Thompson's help, he trained an ANN on the endgame KPK - certainly achieving the objectives of his 2nd-year project but not of course making any breakthroughs in terms of ANNs or KPK. We wrote up the exercise and the paper appeared in the ICCA J v21.4 (1998). As a footnote, we observed that a linear evaluation function could serve as the initial state of an ANN .... and the ANN might develop into, effectively, a non-linear evaluation function on the basis of some training regime. Michael Buro believes - and I am sure he is right - that if 'features' can be named and recognised in positions - Bayesian statistical analysis is more powerful than neural networks (which he says are relatively inefficient). The more one can input to the recognition problem in 'high level human terms', the more efficient your resulting 'black box' is going to be. With ANNs, you put in the least and the ANNs are v slow to learn. There have been one or two more serious ANN-based attempts in chess. See past Adv.Computer.Chess books and Furnkrnatz' bibliography - somewhere on the web - about machine learning. None were particularly successful. Guy
This page took 0.1 seconds to execute
Last modified: Thu, 07 Jul 11 08:48:38 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.