Author: Jay Scott
Date: 10:53:58 03/26/98
Go up one level in this thread
On March 26, 1998 at 13:04:34, Dan Homan wrote: >It >wouldn't be very hard to create a 'profile' for every player >who plays my program and store information about the effectiveness >of certain openings (and other things, as I think of them). I don't know of any performance program that does this. The AI programs that do opponent modeling have been pretty clumsy about it so far, in my view. I think the interesting question is, what information will it do your program the most good to learn? It's easy to learn what openings the opponent plays badly, and aim for them, but that info is very specific. It would be more interesting to learn, for example, that player X folds up under violent attack and player Y can't cope with rook endings. But also harder, of course. My idea is that the more general the information that you learn, the more often you can use it, so the more useful it is. One possibility is to try to categorize your opponents and learn about the categories. The first category split could be human/computer opponent. Another split could be relative tendency to make mistakes, or uncover your program's mistakes, in the opening, middle game, or ending: tabulate the material on the board when there are big score shifts, and use the info in deciding whether to trade down. Another could be playing strength relative to your program; maybe stronger opponents should be treated differently than weaker ones. I'm sure you can think of more. The advantage of the category approach is that as soon as you can assign an opponent to a category, you know that much more about them--and you may even be able to do it by looking at their past games, before meeting the opponent at all. Plus, nothing in the category approach stops you from doing individual opponent learning too. Jay
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.