Author: Don Dailey
Date: 19:21:51 08/26/98
Go up one level in this thread
On August 26, 1998 at 16:03:58, Alessio Iacovoni wrote: >>Why don't you just explain your idea here? If it survives the critiques, you >>will have gained something of more value than the small sum of money you are >>seeking, and if your idea has weaknesses, you would learn about them. > >Ok. Here it is... > >Now, first some facts: > >1. When playing against a chessmachine the "human" player adopts certain >strategies that it feels will make the "machine" loose... > >For example: > >(a) let it go out of it's opening book in the very first stage of the fight (as >Kasparov did at least in one game aginst deep blue if i dont remember wrong). >(b) play closed games >(c) exchange as many pieces as possible >(d) play "calm" and "positional" >(e) attempt an attack on the computer's castled king with a piece sacrifice >(strange but many programs will fall in for it!!) - opening lines and exposing >his majesty. >(f) make positional sacrifices (i.e. "real" ones) > >So basically it seems that the way the program is beaten is through positional >knowledge and a clever use of sacrifices plus etc etc.. i.e. intelligence.. > >But you know all of this.. the solution would be to have 2 engines play the same >game: > >a tactically strong one i.e. Fritz >and a positionally "knowledge based one" > >The tactical engine would be in control all of the time, but all of the moves >would be filted by the "positional one" in such a way as to avoid those >"bluders" which are typical of programs. The filtering could be weighted in >suych a way as to ovveride the positional engine filtering if a very strong >tactical combination is found and viceversa. etc etc.. I'm sorry to report that this is not a new idea. Jonathon Schaeffer used the basic idea in Phoenix, his chess program many years ago. But it is still unclear how workable it is. He used the idea because he felt like it was a better resource allocation decision when using many processors. But integrating the move choice between both engines is much more difficult that you might imagine it to be. For instance, which move do you chose? If the tactical engine says play move A with an even score, and the positional engine says to play move B and be up 1.5 pawns which do you choose? I would guess that one of these programs will be stronger, depending on how they were designed and I would tend to want to choose it's move most or all the time. Another big problem is how strong do you think you can make a program positionally? The real art in writing an evaluation function is choosing the right evaluation features and knowing how to mix them together and with the correct weights. In my opinion this is harder to do than just pumping in more and more knowledge, this very quickly reaches a point of diminishing returns. A point often missed is that a deep searcher will often play better moves, both tactically and positionally. Most people think of these as 2 separate things but they are not. The search is a magnifying glass for the evaluation function, if you improve the search, the positional play will also improve automatically. Can you suggest a scheme for choosing which move you would play under which circumstance so we can talk about it? With Phoenix, the positional program was in control, but the tactical program could "veto" move choices the positional program made if it saw that another move won material. Another problem with your scheme is again, resource allocation. If you have 2 programs, you cripple both of them unless you run them on different processors, but you are implying that this scheme be used for any program. I admit I am pretty skeptical about your idea but I am willing to discuss it further if you want to continue. You can start by being a lot more specific about how you might arrange the move selection process between 2 programs. You also must convince me that two different programs can be written where each one has huge strengths the other lacks, and yet these cannot in any practical way be combined into one program. - 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.