Author: Don Dailey
Date: 16:48:17 06/03/98
Go up one level in this thread
So here is a simple, high level description of Alpha/beta for you non-chess programmers. I'll leave out the nasty details. Pretend we have a position with only two possible moves, lets say they are e6 and Kb5 just for the sake of argument. Now we look at e6, do a search and discover that we win exactly 1 pawn. Great. So now we look at Kb5. Let's say the opponent has 50 different replies to this. The very first one he looks at wins a queen for him. So what do we do now? We KNOW we can play e6 and win a pawn, or we can play Kb5 and lose a queen. We now know looking at the other 49 moves of his is silly, because we already know we can avoid the whole mess by playing e6 and winning a pawn. The only thing we might learn by searching the other 49 moves is that he might win even more than a queen, if we were foolish enough to play Kb5! We "prunned" 49 moves without any side effects or taking any chances. Alpha beta prunning is a free speedup. - Don On June 02, 1998 at 05:03:32, Inmann Werner wrote: >Alpha - Beta is silly? > >Yesterday, I tried the LCTII Test Suite, position LCTFIN01. I was >amazed, that INMICHESS (my program) needed depth 9 to solve the >position. (running pawn) > > >The position: Kc2/Pf2/kc4/Pg4/Pf5/Ph6/pb7/pf7/pg7/ph6 >Best Moves: 1.f6 2.gxf6 3.f4 4.b5 5.g5 6.fxg5 7.fxg5 8.hxg5 9.g6 >....->Queen > >I did not understand, why it did not recognize the running pawn in depth >5. The 3 beating moves afterwards should be made in the quiet search, >and then there is the running pawn! >Nothing at all! I debugged the program. >Very fast it was in depth 5 and then in the quiet search. There it made >the 3 beating moves up to ply 8 and the calculatin function brought back >-844. (more then 8 pawns)!! Super! >But then alpha-beta threw the move away, because it was blacks move, and >black did not want to get this bad rating. This continued until the >normal search was in depth 9, although the black moves are forced. >But think! This all is normal alpha-beta with no programming failure! > >Conclusion: alpha-beta does no moves, where the player who moves get a >bad rating, especially in qiet search, although if the moves are forced. >It thinks, if ist my move, it only can get better. (Something like a >Null-Move thinking). On one side, this makes alpha-beta fast. On the >other side, trivial things get moved behind horizon. >If you try to change this, the program gets worse, because in most >positions, above is right. > >Is above conclusion correct? >Has anybody thought deeper about this? > >Maybe, yesterday the first time, I understood alpha-beta. > >Besides: Fritz 4 also recognized the running Pawn in depth 9! > >Best Greetings > >Werner
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.