home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!usc!wupost!gumby!wmu-coyote!33plato
- From: 33plato@sol.cs.wmich.edu (Larry Plato)
- Subject: Re: NNs in chess or other games
- Message-ID: <1992Sep12.020309.10625@sol.cs.wmich.edu>
- Organization: Western Michigan Univ. Comp. Sci. Dept.
- References: <18psroINNnek@nestroy.wu-wien.ac.at>
- Date: Sat, 12 Sep 1992 02:03:09 GMT
- Lines: 77
-
- bruhn@uxe.wu-wien.ac.at (Peter Bruhn) writes:
-
- >Does anyone of you know something about an application of neural nets
- >in strategic games (esp. chess). What I think about is that a NN could
- >be trained to evaluate a position or some aspects of a position (e.g.
- >the structure of your pawns, how safe is the position of your king,...).
-
- >The idea behind is that evaluating a position is (as I see it) closely
- >related to pattern-recognition. An experienced human chess player is
- >able to identify a position as weak, even though there are no tactical
- >dangers to perceive and he/she has not seen the position before. So how
- >does he/she know? I think it is due to pattern recognition: He/She has
- >seen similiar positions before or has made the experience that a certain
- >pawn structure is a weak point. Once a chess master (I don't remember his
- >name) was asked how many moves he prefigures in advance and he answered
- >"normally: none". Even though this is an exageration, of course, his
- >answer shows that it is more important to be able to evaluate a position
- >than to prefigure a lot of moves! Interestingly todays chess programs
- >take the other way: the use brute force! And this is the reason why
- >computers are good in chess, but not in games like Go. In Go there are
- >more possibilities and brute force will not work so well. So what I
- >think is that we should put more emphasis on the evaluation of positions
- >(of a game) rather than on brute force. And maybe NNs could help !?
-
- >Any comments are appreciated!
-
- >Peter (bruhn@uxe.wu-wien.ac.at)
-
- I wholeheartedly agree. I am an undergrad at Western Michigan U and
- a senior in the CS program taking Neural Nets. I think a three tiered approach
- would be best. As you mentioned, a complete search of the game space tree
- in prohibitive. So what we do is this...
-
- a) a neural net analyses the board and produces a list of 'insightful moves'
- that are worthy of consideration.
- b) an expert system could analyse the board based on more traditional
- factors, such as fiancetti, pawn location, freedom of movement of the
- major pieces and territory controlled.
- c) a brutre force approach could be applied to the promising moves.
-
- to further analyse my ramblings imagine this . . .
-
- A traditional program does b) quite well, and a certain amount of a.
- this is referred to as pruning, because you 'prune' the game space tree.
- The only thing we are doing by adding the neural net is allowing the
- 'rouge' element to come to the fore. The net might see a promising
- pattern that a more traditional view would miss. This is sort of like
- a local minima problem. If you start pruning the tree by some objective
- function, then you will find a local minima (eg a good solution) but you may
- or may not find the 'global minima' which is to say the best solution.
-
- A couple of old jokes here, the first is have you heard about the US Army
- chess manual? chapter 1 is white kings pawn to kings pawn 1 chapter 2 are the choices for white kings pawn to king pawn 2 etc ad nauseum.
- The other is the chess master who draws black and is so good the game goes
- white queens pawn to queen's pawn 2 and black resigns.
-
- The 2 problems I see immediately are
-
- 1) how to choose a network configuration?
-
- 2) how to run backprop on a net with no immediately obvious objective function
-
- I am working on a idea to represent nets so I can run GA's on them but if it
- goes like most of my projects I would not expect results this decade.
- as for 2) I would use famous games as a data set. Start with the move before
- checkmate, train the net on the 2nd to last move of a hundred famous games,
- when it can extend the game some of the time then go to n-2 moves etc. I would
- avoid porting in the expert system, brute force crap until the net got pretty
- good, we want a 'rouge' element not a random one.
- best of luck
- Larry Plato
-
- --
- --------------------------------------------------------------------------------
- There are many ways to die, some are less pleasant than others.
- 33plato@sol.cs.wmich.edu
- --------------------------------------------------------------------------------
-