Author: Matt Frank
Date: 17:27:03 03/09/99
Go up one level in this thread
On March 09, 1999 at 10:01:47, Marc-Philippe HUGET wrote: >>I'd say that depth of search, evaluation, and if the evaluation proceded to >>fininsh; final result are important in the learning function. I would suggest >>that the learning should be focused on high level games and if necessary the >>analyses used for learning should be constrained to only high level opponents. >>Hope this helps. >> >>Regards, >>Matt Frank > > >Yes, but it is not very easy to know if your opponent is good or not. What do >you think of book learning where we have position, score and best move for this >position. While this best move is not refuted, we store this move, so the >program only plays this move for this position. > >mph Well the quality of the games has to be a process of proper selection by the "teacher," YOU!!! Yes, saving positions that were followed by significant changes in the evaluation function should be considered by the program as the potentially most salient positions (whether positive or negative). In any event, such evaluation changes should result in the program endeavoring to obtain the most favorable position previously stored either reachable through a transposition of moves, or through a forced process. And, a good learning process would alter the evaluation of previously considered drawn positions. By way of explaining, if a program encountered 5 opening variations, all assumed to provide even chances for White and Black (according to present theory), yet found that variation 5 was the variation that provided the clearest advantage for either side (say White for this example) then White would most love to reach that position and Black would most eagerly avoid such a position (i.e., the position reached by that variation). Of course the weighting of moves and other factors included in the evaluation of positions requires careful thought to the heuristics applied in the program, and that topic is a lengthy and complicated topic indeed. However, simply put, any good learning program would seriously weigh a significant change in evaluation functions very seriously. Regards, Matt Frank
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.