Author: Sune Fischer
Date: 15:24:05 03/10/04
Go up one level in this thread
>That's a good point. There would certainly be some "structural error" based on >the limitations of the class of functions we decided to train. But given the >function class (which we've assumed is the same for both hand tuning or machine >tuning), my main point is that an automatic approach (attempting global >optimization of the parameters) could get the most performance out of the chosen >function class. There is a small problem in that we don't have a good cost function. It would be 'easy' to automaticly tune the evaluation parameters if we had a few billion positions with accurate score values. Unfortunately there are only three accurate score values (won,drawn,lost), if we want inbetween scores we have to do manual estimates. As you can see that would have little to do with machine tuning, one could just as well try and estimate the parameters directly. -S. >Michael
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.