Author: Don Dailey
Date: 15:20:19 06/10/98
Go up one level in this thread
On June 10, 1998 at 17:59:22, Fernando Villegas wrote: >On June 10, 1998 at 07:26:02, Mark Taylor wrote: > >>On June 10, 1998 at 05:58:53, Carlos Adan Bonilla wrote: >> >>>...the computer may select a >>>complicated line rather than an "easy-to-see-for-humans" line, even if >>>that line is not so good. >>>Also, in lost positions, the computer will select the lines in which >>>there is an opponent move that is a mistake and could turn the game back >>>to a drawing position (or winning). >> >>I have had similar ideas & believe that statistically this sort of >>strategy would gain better results (esp. against human & weaker computer >>opponents). The strategy does not mean playing weaker moves, when there >>is a better move available. It means playing a move that gives the >>opponent a choice of one good move out of ten bad moves, rather than >>playing a move where the opponent has one bad move & ten good ones, even >>though an alpha-beta search might return identical scores for the two >>moves. > > >I have thought of this same idea and even I posted here speculating that >a kind of device like that could be one of the secrets of CSTAL. >Precisely I said that the CSTAL style suggested that he did not use the >usual alfa beta porunning techniques, but something different, something >where the main purpose is creating pressure, not just to suppose the >adversary will do the very best move >Fernando > >I have thought about how this kind information could be gathered in an >>A/B search without slowing things up - it would mean searching the very >>sub-trees that A/B currently ignores. Maybe this could be gathered in a >>search on the programs' turn when TOOT has predicted the opponents move >>(& the programs' next move has already been computed). >> >>Any ideas anyone? I tend to doubt the idea it does not do alpha/beta pruning. It's style could perhaps be explained simply by aggressive evaluation. Thorsten tried to explain it once to me but it was in such high level language, there were no useful details (which perhaps was the intent!) I got a vague impression it was pretty selective but this doesn't explain an aggressive style. When playing Cilkchess against it, I noticed it's scores were often way below or above a pawn even in positions that seems relatively even to me. I think it might be possible to create such a program by making key evaluation terms very aggressive. I imagine getting it right takes a lot of tinkering. It often plays completely unsound moves, but gets away with it. Its strength seems to be based more on "playing the opponent" than technically correct chess. But this is truly impressive since I view this as the more difficult task. It truly does play like "Tal", knowing which moves give it winning chances despite the "theoretical correctness" of the move. - 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.