Author: Simon Waters
Date: 07:18:01 11/09/00
Go up one level in this thread
I actually meant that it may have extended it's opening book so it won't play the same line as the program you have at home, so your preparation might be less effective than expected. I guess in that sense it is no different than preparing an opening rep against another GM - you'd have to keep abreast of recent games. Just the machines do this somewhat faster, if with less understanding, than the average(Oxymoron?) GM. Deep Blue has a tuning utility (in the public domain I believe) so it can retune it weighting on the basis of past games. This is similar I suspect to the old Samuel's Checker program that used it's own search to improve it's evaluation function. Bob muttered in his interview (See the archive) that Crafty had some middlegame learning ability - but I haven't seen it in the code so maybe he dropped it? Or maybe it is just hashing some positions and scores. To get real improvement in the middlegame the machines would probably have to do more than just "experience the middlegame" - Samuel showed a simple way to use the machines own analysis to optimise weightings (Few people seem to do it as it is actually tough to do if the terms in the eval function are at all complex, and if it is complex you are likely to find stable but suboptimal solutions). I suspect you could program a genetic approach to chess evaluation functions, by varying weightings and comparing performance - but it is slow and achieves a similar effect to what Samuel suggests, but require less knowledge of statistics and numerical modelling to implement. Until you can get an evaluation function to posit new potential terms in the eval function (That have a vague hope of success - I suspect random terms in the evaluation functions would be of limited value) they won't really improve themselves much. The problem with this approach is you need a formulation of the evaluation functions that lends itself to easy mutations that are likely to be helpful terms, or that can be used to describe existing games. Anyway - given me an idea for a way of improving the program I'm working on.
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.