Author: Don Beal
Date: 09:32:18 10/02/98
Go up one level in this thread
On October 01, 1998 at 20:28:55, jonathan Baxter wrote: >Currently I am working on an improved version of KnightCap, with better >search and a much *simpler* evaluation function. Interesting preliminary results >show that with parameter learning and *just* piece/square tables we can get >a rating of around 2500 on ICC. This is without any opening learning (the Also Blitz play, and blitz rating? >Another interesting observation is that the material values, although >they start out at 1 4 4 6 12, evolve towards approximately 1.25 3.9 3.9 6 12 >which is much closer to the traditional 1 3 3 5 9 (if you rescale so a pawn is >1). We found the material values converged to something close to traditional ratios whatever our starting values. However, our absolute values vary with other factors, such as what other weights are being learnt. >For those who are interested, I have included all the final piece/square tables >below, after learning for a few hundred games against crafty. Our initial experiments were with a material-only program. We then learnt piece-square tables, and saw elementary traditional knowledge appearing, e.g. advance pawns, centralise knights, get pieces off their starting squares, invade the enemy side of the board, etc. I'll try to get values to post (I'm currently away from home base, visiting MIT, and not everything is online.) We've been mainly interested in methods that learn from no starting knowledge and learn from self-play (such methods may be useful in future for tasks where no human knowledge exists, or where the computer is expected to go beyond human abilities), so the emphasis is slightly different from Knightcap. It's fun to run these methods, and see "knowledge" appearing in the weights. My first reaction was "wow, it's really learnt some chess knowledge - by itself". It's only weight-tuning, of course, but it still impresses me that the weights match (roughly) elementary chess advice given to beginners. > Don Beal wrote: >> The method works for any weights, not just piece values, provided >>that the evaluation consists of a sum of term*weight components. > >This isn't true. You can use the method for any evaluation function provided >it is a differentiable function of its weights. This includes linear evaluation >functions, but is not restricted to them. >Cheers, > >Jonathan Baxter Erm, I think you mean: not only is Don's statement true, but the method can also be applied to a wider range of evaluation functions, provided it is differentiable with respect to each weight. My statement was correct. Your point is correct too. Best regards, Don Beal. BTW, we applied the same method to Shogi, and learnt piece values there. That's directly useful, because Shogi doesn't have a standardised set of values that programmers can pick up and use.
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.