Author: Don Dailey
Date: 15:27:06 01/12/98
Go up one level in this thread
On January 12, 1998 at 17:43:30, Stuart Cracraft wrote: > >What are the experiences of this group with automated tuning? > >I'm thinking of two specifically: > > - temporal difference learning (Tesauro-style) or best-fit > (Nowatzyk/Deep-Blue style) based on master games > > - survival-of-fitest round-robin matches between same program > with slightly different coefficients or weights for evaluation > function > >--Stuart Hi Stuart, I played around with genetic algorithms for this. It was a heck of a lot of fun but nothing came out of it. I think the problem has to do with the incredible slowness of the fitness function where you have to play games to measure fitness. 1 or 2 games is a very poor fitness function. But nevertheless, I did come up with interesting results. The process eventually returned realistic values for the pieces which gave me some confidence that at least in principle the idea might be workable. John Stanback also played around with this stuff. I did something similar with master game databases with Larry Kaufman. We looked at hundreds of material signatures and computed values for all the pieces using the genetic algorithm. Our results were based on which piece combinations tended to win a lose. I remember little except that the numbers were very normal looking, even more so than the prior experiment. The one odd thing was that the bishop pair got a very high value. But our program over the years has kept increasing the value of the bishop pair, I don't think the number was very far off now. It came to almost half a pawn. We believe the right value is close to this. But of course there are numerous factors involved. I remember a master once telling me that 2 bishops were more than a pawn better than 2 knights so this value may not be ridiculous. We also believe the bishop and knight (not considering the bishop pair) is about the same with perhaps a very slight nod to the bishop. Also there are conditions as we all know where a knight is much better and so on but I'm talking about the "average" case with nothing else known. We have developed a sort of theory of piece cooperation you may be interested in hearing. The main point is that similar pieces do not help as much, in other words 1 knight is good, 2 knights are not quite twice as good. But 2 bishops are completely different pieces because they are on different color squares. The second black squared bishop for instance would be worth very little compared to the first. Queens share some powers with bishops and rooks so queens might be better with knights. Our program gives a rook pair penalty, a knight pair penalty, a bishop pair bonus and others. Most of these numbers are quite small and probably affect things very rarely. Most people are surprised at the rook pair penalty but I've had a couple Grandmasters after scratching their heads a bit say it made a lot of sense. One important note: No matter what numbers you use they must make a lot of sense in relation to other things. If you have very conservative evaluation you would be foolish to use 1/2 pawn for bishop pair bonus because your program would be "heel bent" to hang on to the bishop pair and seriously compromise it's position to do this. - Don
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.