Author: Robert Hyatt
Date: 18:57:07 07/12/00
Go up one level in this thread
On July 12, 2000 at 02:40:58, Tony Werten wrote: >On July 11, 2000 at 15:19:29, Robert Hyatt wrote: > >>On July 11, 2000 at 05:20:07, Tony Werten wrote: >> >>>On July 11, 2000 at 04:23:25, Dann Corbit wrote: >>> >>>>On July 07, 2000 at 06:26:46, Adrien Regimbald wrote: >>>> >>> >>>>I think a hybrid approach is best. A hand tuned book by top experts is going to >>>>be better than an automatic book. However, it will cost millions of dollars to >>>>hand evaluate many millions of positions. In short, it's not gonna happen any >>>>time soon. So you are left with a small (perhaps nonexistant -- depending upon >>>>the size of your wallet) book that is written by hand, and a much larger >>>>automatic book. I think that even the hand-tuned book might possibly benefit by >>>>using the right mathematical approach. >>> >>>I agree with that hybrid. If you add a flag to every position indicating if it's >>>automatic- of hand generated, and mini-max it to two scores, one for only >>>handevaluated positions and for one all positions, you have best of both if you >>>let your engine prefer the first. >>> >>>Then, whenever you see your engine ending in a position it doesn't understand >>>you just hand-adjust the score for that position, set the flag, minimax, and >>>both your scores will benefit from it. >>> >>>Some goes if you want to add some different variations. Point is that at least >>>you have something until you have done that. >>> >>>Tony >> >> >>Minimaxing simply doesn't work. I did this in Cray Blitz, and in early Crafty >>books. It simply allows scores to be backed up way too far... with no regard >>for other moves that are not included. You end up playing down lines that are >>going to get you murdered. Because one key move is not present and so didn't >>influence the minimax score. >> >>A cap score used "at the point the position is reached" is perfectly reasonable, >>as all that is missed is whatever the program producing the analysis missed. >>But when you start minimaxing, you make the errors multiply... > >I see this can happen if you only evaluate the leaves and then minimax. But if >you also have evaluation inside the tree, you should have found great moves that >weren't played. So if you reach a point where the pv suggests a move that wasn't >in the cap data, use the score for that move as if it was. > >Tony I don't see how to do that. The CAP data is composed of _very_ deep searches. While building a book from millions of games, there is _no way_ to do searching at each point in the tree. Even at 1 second per search, 2 million games has something like 160 million moves. Doing 160M 1 second searches will take a _long_ time. And 1 second is _not_ reliable for choosing book moves.
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.