Author: Robert Hyatt
Date: 05:10:17 05/20/98
Go up one level in this thread
On May 20, 1998 at 00:01:45, Christophe Theron wrote: >On May 18, 1998 at 21:18:24, Dave Gomboc wrote: > >>On May 18, 1998 at 18:12:53, Ren Wu wrote: >> >>> >>>On May 18, 1998 at 17:14:10, Carsten Kossendey wrote: >>> >>> >>>>First of all, your output does in no way state which scores are bounds >>>>and which are exact, so everyone will assume they are *all* exact. >>> >>>First of all, you != everyone. So you is you, please don't confuse >>>youself to everyone, which is a big mistake i think. >>> >>>>Secondly the point of the position was that Bxa8 is a *really* bad move, >>>>and a score of <= +3.81 is not exactly suggesting this. Crafty's +0.56 >>>>eval (contrary to +3.16 for b6) gets a lot closer to the point, for >>>>example. >>> >>>When you know a move no good, it is little interets to find out exactly >>>how bad it is, especially if you found another move is better. >>> >>>I will not answer any follow up on this thread, because i found little >>>info it contains. What i want say at my orignal post is that a program >>>with little king safety knowledge can find the correct move at 9ply, and >>>i don't understand why other top program have difficulties. >>> >>>I have little interests to prove my program is stronger than others, it >>>is not at the moment. I am more interested in why those programs have >>>diffculties. >>> >>>And i am happy that my program did good job at this position. >>> >>>Ren (renw@iname.com) >> >>I have a hard time believing that the +3.81 for Bxa8 was an upper bound, >>not an exact score. It was the first move searched at that ply level. >>Unless you're doing something strange, that move will be resolved to a >>score before alternatives are looked at. >> >>Dave Gomboc > >No Dave. You don't know what he is doing is his program. I don't either. > >But a standard implementation for alpha beta with aspiration could >behave this way. That is, if the first move in the list fails low (has a >score below alpha), the program doesn't try to compute the exact score >and PV for this move, and goes directly to the others moves without >showing any PV. > >So he may choose, at this time, to display the first move with a >(default) score equal to alpha (the lower bound), just to show what's >happening, or display nothing at all. > this is possible, but where does the PV come from? IE to fail low at the root, there was a "good move" found at ply=2, but not necessarily the "best" move. And at ply=3 you have a problem, in that *every* move was refuted, so there's nothing to tell you which move was best... same is true for every odd ply move, and for ever even ply move, you only have a "good" move and not a "best" move... >If no other move is able to bring the score into the alpha beta window, >the search starts again with a lower window. > >For example, in case of a fail low, Chess Tiger displays nothing after >the first move has been searched. If another move manages to bring a >score inside the window, this move and PV is displayed, as if it was the >first to have been searched (it may look strange when you follow the >computations, but it's OK). > >If no move is able to bring the score above alpha, Tiger displays >"Oops..." (self explanatory I think) and starts the search again with a >lower window. You can easily imagine that I hate to see this message... >:) > >There are others ways to do it, that's why you shouldn't draw >conclusions just by seeing a program's log file. > >The only thing that can be trusted (I believe) is the first move of the >PV. Everything else (rest of the PV, score) can be wrong in a particular >program. For example, it seems that Deep Blue's PV cannot be trusted >(and it's going to take a century or 2 to explain that it can happen, >and that it doesn't prove any cheating on DB side). > > > Christophe
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.