Author: Vincent Diepeveen
Date: 09:16:01 05/21/02
Go up one level in this thread
On May 20, 2002 at 08:23:35, Eric Baum wrote: >How much do modern programs benefit from >developments beyond alpha-beta search +quiesence >search? So, if you did the same depth search, >same quiesence search, same opening book, >same endgame tables, but replaced the evaluation >function with something primitive-- say material >and not much else-- how many rating points would you >lose? Programs especially advanced in evaluation knowledge and even more in weakest chain. Taking into account weakest chain, for DIEP i would say 600 points. In past times i could easily trick it in positions it knew shit from, Its weakest chain now is really *not* so weak. >My recollection is that one of the Deep Thought thesis >showed a minimal gain for Deep Thought from >extensive training of evaluation function-- That's a complete question to me. They must have put complete nonsense into the evaluation function by default in order to win from automatic training. Automatic training simply *does* not improve an evaluation function. Only makes it weaker instead. >it gained some tens of rating points, but >less than it would have gained >from a ply of additional search. Has that changed? It is important to search 6 moves, or nominal said: 12 ply. Note that this is not 12 ply for all programs, because program A counts plies different than program B. For junior 12 plies is like 3*12 = 36, whereas for fritz it looks more like 20 = 12 positional plies. For programs like Yace, DIEP there 12 ply = 12 ply. For crafty which is incredible weak in the last few plies of it, i would say 14 ply. A few years ago no one got to that 12 ply. Even deep blue hardly got it. It got 11 to 12 ply, though it compensated more than enough by tactics. Getting another ply simply doesn't bring much better play. Improving evaluation and openingsbook is much more important. Right now in many strong programs the weakest chain is book; some commercial programs are so well prepared in opening, that i find it shocking to see games where one side clearly is going to win the game, and of course it takes only 30 moves for the programmer to realize that his thing is losing, as he doesn't know about the good strategical job which is done by the book makers of his opponent.
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.