Author: Robert Hyatt
Date: 09:52:44 01/14/04
Go up one level in this thread
On January 14, 2004 at 08:02:03, James T. Walker wrote: >On January 14, 2004 at 07:08:50, Tord Romstad wrote: > >>The day before I released Gothmog 0.4.5, I played a blitz match (4 minutes/game, >>with 1 second increment) between my engine and Phalanx XXII on my PowerBook >>G4 550 MHz. Gothmog narrowly won the match; the final score was 52.5-47.5. >> >>Yesterday, I started a new match between the same two engines. The only >>difference was that this match was played on a PIV 2.4 GHz, and that learning >>was disabled for Phalanx in the second match (my own engine doesn't learn). >>The result: 65-35 for Gothmog. >> >>Is this just a statistical fluctuation, or is learning really that effective? >>Or perhaps Phalanx (a very old engine) simply doesn't play well on fast >>hardware? >> >>Tord > >I think the answer would lie in how many games were repeated. (I assume you are >talking about book learning) >I have done a test on the effectiveness of book learning by playing the same >program vs itself after one had "learned" from 1500 games and the other had no >previous learning. There was almost not difference in performance. I belive >the only real benefit of book learning is to prevent repeating the same loss >over and over. >Jim Learning is really (in general) an opponent-based issue. IE I might beat program X with a particular opening every time, but lose to program Y with the same opening. Reasonable learning will handle this just fine, the first time I play Y I try the opening, the next game I won't as I just lost. Learning really will influence a match between two programs, because it makes the "loser" do something _every_ game to try to avoid that same loss again. If the loser doesn't learn, then it is going to get smoked. Learning _does_ apply across opponents for some cases. IE some particular openings are just bad for computers, period, and this learning will carry from opponent to opponent. But much of what is learned is "opponent specific".
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.