Author: Dave Gomboc
Date: 11:23:29 03/10/04
Go up one level in this thread
On March 09, 2004 at 16:05:15, Gian-Carlo Pascutto wrote: >Yet no top program does this, and they had a human correct it >afterwards in Deep Blue. The conclusion should be obvious. Is that so? >If you can develop a *top level* evaluation function, better than >good human tuning, solely on learning from a GM games database, >you deserve an award. Nobody has succeeded before. Jonathan Schaeffer learned weights in Checkers [Chinook] without even using a human games database (he used TD learning). The weights he tuned score 50% against his hand-tuned code. I learned weights in Chess [Crafty] using 32k positions, hill-climbing an ordinal correlation measure. It too scores 50% against the hand-tuned code. Given Deep Sjeng's source code, I could zero its evaluation function weights, and learn them from GM games to score 50% against the weights you have right now too. (I'd need to make some performance improvements to my current tuner to tune table-lookup-based terms efficiently, but that's an implementation issue, not a research issue.) Weight tuning is no longer the issue. Selecting the features that will be evaluated is the issue! Dave
This page took 0.01 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.