Author: Dan Andersson
Date: 09:19:36 12/29/01
Go up one level in this thread
One problem is that you are not guaranteed to stop. There might be other positions in the search that gain more from the changes in evaluation. I know that's not exactly true, since a piece square eval might put arbitarily large values in the tables. Thus creating a 'magic' evaluating function. That can be avoided if you do a check after each evaluation function change. So that you do not break the result of the other positions tuned for. That means that in the best case you would have to search O(n^2) for n test positions (And n needs to be large to avoid a meta-'magic' evaluator). And in the worst case you would never stop until you added new higher order evaluation terms in your program. Providing the needed knowledge to solve certain positions. But you would have to avoid a meta-'magic' evaluator, keeping enough information from each test position to differentiate them and indexing a 'magic' number. The optimisation domain seems more amenable Genetic Algoritms and Genetic Programming than simple differntial learning, due to the immense search space. MvH Dan Andersson
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.