Author: Amir Ban
Date: 04:31:38 10/18/00
Go up one level in this thread
On October 17, 2000 at 00:53:58, Ratko V Tomic wrote: >> The only thing I was pointing out is that if a program evaluates >> something as +3, when material is even, then it will evaluate that >> same position as even, when it is a piece down. > >That doesn't mean at all it would give a piece in that position. >Only a simple linear/additive form of the evaluation function would >have such side-effect. But a nonlinear combination of separate evaluation >components (as exemplified by Samuels' checkers program) need not >have such drawback. > >While Gambit Tiger might not go as far to use as general mapping >as Samuels program did (since the chess parameter space is vastly >larger, in extent & number of dimensions, than that of checkers, and >would thus take too long to tune), there are great many intermediate >levels of non-linearity, some of which might work better and be practically >tunable in chess. For example, the sum of squares of properly chosen >term deviations would give much greater weight to the term which stands >out in a given position, producing an effect of program concentrating on >that aspect/feature of the position. That is in fact much more humanlike >way of analyzing a position. A good player knows what is the most relevant >in a position and spends his computational resources on the relevant >aspects, ignoring the irrelevant ones. > >The Botvinnik's scheme takes this concept (of over-emphasis of relevant) >further by explicitly guiding multiple alpha-beta searches from the same >position, but using different objectives (and evaluation functions) each >time. In that case you would see great jumps in evaluation, which >merely reflects different perspectives one might take in a given position. > >It is natural that if you analyze a position looking for, say, a king >attack, then analyze the same position looking for a better pawn structure >for endgame, you will have widely varying estimates, e.g. if king attack >doesn't seem useful, you may get a very low score in that search (due >to weakening/sacrifices needed to get the attack off the ground). At the >same time the pawn structure oriented search may give an even position. > >On the other hand, if the king attack search looks promissing, triggering >many criteria for a strong (but beyond horizon) attack, you may get a high >evaluation, well beyond the material balance on the board. For example >you could get +3 value, but that doesn't mean you would get 0 value if >you take a piece off the board (e.g. that piece is contributing to >the king attack or defending against the break on the other side etc). >Without a piece, there would likely not be any triggering of the high >score for the king attack, any you might simply get -3 score. > >Taking one more step in abstracting the meaning of such high evaluation >swings (the first two levels being nonlinearity and different perspectives), >one can view the large score (well beyond the simple additive variation) as a >reflection of a much longer term estimate. This longer term estimate comes >from specialized evaluations tuned for a given type of position, which >may not be available or usable in all positions. But when it is judged >usable by the program, it will give a value much closer to what the >regular evaluation may see 20 or 30 plies later. E.g. in getting >ready for a king side attack, no checkmate or large material gain is seen >in the nodes examined, but the specialized long term estimators indicate >that in such position the expected gain may be a piece by the time attack >is completed, some 20-30 plies later. The large score variation is thus a >kind of far extrapolation reaching much closer to the final game values >(which could be, say, +99 for a win) than what regular evaluation offers. > >Obviously, the trick is to come up with such long term evaluations. But, >observing human play, we know these do exist in some implicit form in the >mind of strong players. So, since they do exist, it isn't beyond possible that >someone (Christophe T., Chris W., David K., Marty H., ...) may have teased out >some of that knowledge. I disagree. There is no need to go so far as to assume the need for multiple points of view of the position and so multiple evaluation functions, non-linear terms etc. You tacitly assume Bob's argument: That being a piece down cannot be justified by position terms. This is certainly true for a wishy-washy, inexact evaluation, but I don't believe it to be true when the evaluation becomes better. I have enough test games where I thought my program was commiting quick suicide, getting as much as a rook down in an open position, but got away with it, to convince me that this is possible. I don't see any reason to focus on some aspect of the game and ignore the others, i.e. because there's a king attack going on, there's no need to discount pawn weaknesses or such stuff. It's just a question of giving the correct weight to everything. In other words, good evaluation is all that is needed, and the magic will follow. Getting it is quite a trick in itself, of course, and the first step is to understand what is meant by "good evaluation". Amir
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.