Author: Tony Werten
Date: 22:31:39 07/14/05
Go up one level in this thread
On July 14, 2005 at 19:12:44, Mathieu Pagé wrote: >On July 14, 2005 at 16:03:19, Reinhard Scharnagl wrote: > >>On July 14, 2005 at 15:29:31, Mathieu Pagé wrote: >> >>>On July 14, 2005 at 15:14:35, Reinhard Scharnagl wrote: >>> >>>>Hi Mathieu, >>>> >>>>I have made some thoughts on it. It tends to say good bye to negamax approaches. >>>>Because the same position will be evaluated differently from personalized points >>>>of view. >> >>>I do not see why we could not use a negamax approach with a opponent based >>>learning. >>> >>>The learning i'm talking about is book learning and weight learning. Thoses two >>>techniques have already been used with succes in conjuction with negamax. >>> >>>Maybe you are thinking of some more advandced modeling technique. If it is the >>>case i'd appreciate if you share them with us. >> >>Hi Mathieu, >> >>it will sound to hard, but such an approch is contradicting and will fail. >> >>I see the problem from a very different point of view. Chess is regarded to be >>a zero-sum-game. But this is only true, having full information at hand. >>Inventing detail evaluation functions supporting an engine with values distinct >>from +1, 0, -1 is already proving, that chess could obviously not be handled as >>a zero-sum-game. But the negamax approach is only working using that assumption. >> >>Having different evaluation models for engine personalities will mutate engines >>from evaluation models into prediction models, which might be more effective, >>but establishes the need for navigating through trees with pairs of evaluations. >> >>Reinhard. > >Hi Reinhard, > >Obviously we missunderstand each other. While playing, an engine using my idea >of opponent-modeling will do the same thing as if it was a normal engine. The >difference is that it will use a different book and a different evaluation >function. > >I can not see why it could not work. Opponent modelling assumes using 2 evaluation functions. Yours (the "correct" one ) and your opponents ( the "wrong" one). The point is maximizing your score (the best position) while also maximizing your opponents one (the most likely continuation for your opponent). Alpha beta (negamax etc) will fail to do this since it tries to minimize one, wich you need to "correct" your maximizing score. When using only 1 evaluation function you will not be able to notice wether the score comes from the weaknesses of your opponents evaluation. Tony > >Mathieu Pagé
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.