Author: Robert Hyatt
Date: 13:26:04 05/21/02
Go up one level in this thread
On May 20, 2002 at 14:47:24, Eric Baum wrote: > >OK then: >(1) How much have computer programs benefitted from additional >features? Remove all additional features from the top programs >except material/piece-square table, and how many rating points would you lose? >I'm guessing less than 100, but do you have another estimate? No idea. For Crafty, all improvements over the last 3+ years have been _exclusively_ in the evaluation. I haven't changed the search at all... > >(2) Are there any programs with significant ability to discover new >features, or are essentially all the features programmed in by hand. >If you believe there are programs that discover useful new features, >how many rating points do you think they have gained? >And can you give me some idea of what type of algorithm was used? You are talking about "learning" as humans do it (discover new features). I don't know of _any_ program that does this. Some use pre-defined features, but twaddle with the weights associated with them. But that is very crude in comparison to human learning. > >Also, for comparison, does anybody have a recent estimate of rating >point gain per additional ply of search? 50-70 seems to be current value... has been for years too... > >(3) Also, am I right in thinking that modern programs are still more or >less doing alpha-beta with quiessence search, or has there been real >progress on context dependent >forward pruning, leading to substantial rating points gains? There is some forward pruning going on. From null-move that many use, to real forward pruning as defined by Shannon 50 years ago... > > >On May 20, 2002 at 12:54:48, Robert Hyatt wrote: > >>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? >>> >>>My recollection is that one of the Deep Thought thesis >>>showed a minimal gain for Deep Thought from >>>extensive training of evaluation function-- >>>it gained some tens of rating points, but >>>less than it would have gained >>>from a ply of additional search. Has that changed? >> >> >>You are mixing apples and oranges: >> >>apples: which evaluation features does your program recognize? >> >>oranges: what is the _weight_ you assign for each feature you recognize? >> >>Those are two different things. The deep thought paper addressed only the >>oranges issue. They had a reasonable set of features, and they set about >>trying to find the optimal value for each feature to produce the best play. >> >>Adding _new_ evaluation features would be a completely different thing, >>of course...
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.