Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!swrinde!ucsd!sdcc6!zaius!pluto From: pluto@zaius.ucsd.edu (Mark Plutowski) Newsgroups: comp.ai.neural-nets Subject: Re: Chip vs Chess Master Keywords: Games, chess Message-ID: Date: 27 Mar 91 23:47:46 GMT References: <31403@usc> Sender: news@sdcc6.ucsd.edu Distribution: comp Lines: 80 mamisra@pollux.usc.edu (Manavendra Misra) writes: >Last night, I happened to see a Nova episode which talked about a match-up >between Gary Kasparov (the current Chess world champion) and Deep Thought >(AI software developed at CMU and IBM which is currently the best chess >playing program). Deep Thought essentially carries out a brute force search >of the game tree in deciding its moves. There was lots of discussion of the >standard "computers taking over the world" arguments [...] much to the relief >of most of the people gathered for the match, Kasparov won it. >I was wondering if there was research going into developing "neural" >algorithms for playing games like Chess. [...] >Also, I'd like to hear what others thought of the program. >Manav. It was of course, an excellent production, as are most Nova programs (Donate to PBS!!! You might be on it yourself someday! =-> Least surprising was the hostile reaction of the audience to the computer - and the lack of depth of the reporters questions to Deep Thought's developers. The developers were quite careful to put the right spin on their interpretation of Deep Thought's achievements. I think we ought to expect this type of reaction to intelligent computers, at least until the next generation of hackers grows up. Most interesting was the introspective process Kasparov and the other grand masters tried to describe as they fielded questions about how they evaluate a tough move. They did alot of emoting here, saying things like, "it feels right," "I gain energy from my opponent, whereas the computer is a black hole," or, that they draw upon "inspiration" to break down the opponents "spirit". Sounds like a black box process to me with little conscious symbolic explanation capability. A quite famous (and successful) neural-network game player is the NeuroGammon implementation of backgammon. (Look in 1989 or 1990 NIPs proceedings. I am unsure about the author's name but Tesauro (sp?) comes to mind.) The implementation was most interesting to me due to the way it was trained. (That's right, trained, not programmed, as was Deep Thought.) Out of several ways of training the network to evaluate a board position, the one that worked the best operated by giving the network two board positions, and then (in essence) telling it which one was better. From this relativist style of training, (i.e., without reference to a concrete reference point) the network was able to formulate an internal model which provided a score for any given board position. Said again, it learns to create an internal point of reference, such that it can evaluate (score) a given board position, even though the teacher did not give it these scores during learning. I suspect that much of the task of learning to evaluate a chess board position could be reduced to the process of function approximation, something networks do well. However, scoring a chess board position would seem to be much more context-sensitive than scoring a backgammon board position. Could a network learn to predict trends in the style of play (aggressive, defensive, counterattacking, plodding) of an opponent based upon a sequence of moves? This is just time-series prediction. hmmmmm..... Or, could it learn to place more or less importance upon branches of the game tree based upon how the opponents forces are coalescing on the board? This could provide the computer with a crude ability to adapt to an opponent's strategy. Of course, the real goal is for the computer to develop its own strategy. Get your network to do THAT! IMHO, it's eye-opening that brute force does so well given the high branching factor of the game tree. More elegant approaches to game-tree search would result if future Computer Chess Olympiads placed the computers into categories according to resources used (flops, memory, hardware, power consumption) giving each computer a handicap according to how much brute force it relies on. -=-= M.E. Plutowski, pluto%cs@ucsd.edu UCSD, Computer Science and Engineering 0114 9500 Gilman Drive La Jolla, California 92093-0114