Author: Robert Hyatt
Date: 20:08:40 08/18/00
Go up one level in this thread
On August 18, 2000 at 18:03:28, Dann Corbit wrote: >There has been some discussion about the use of test suites and their usefulness >for program improvements. Personally, I plan to use them extensively to fiddle >with "bean-counter" in order to try to get things right. > >Since there are so many variations that are possible with hundreds of >parameters, I was planning to use gradient search error minimizations with the >evaluation function to try and find an optimal value for all the parameters that >solves a test set of perhaps 5000 carefully verified positions. (Iteration >would be so expensive it would be impossible to use it). The experiment would >be repeated at different time controls, as perhaps some parameters are also a >function of time! > >Now, I am wondering (since at least one of the world's best chess programmers >does not use them at all) if it is such a good idea. So, I am wondering, if you >do not use test positions to tune your evaluation parameters, how on earth do >you choose suitable values for each positional, tactical, and material >parameter? What are the alternatives? Why are the alternatives better? If >test positions were used in the past and abandoned, what prompted the change of >heart? If test positions have *never* been tried, how is it known that they >won't be useful? First, 99% of the test positions that are around are _tactical_. Yes, many of them can be solved with positional tricks, but is that reasonable? IMHO, no. I would _love_ to have a good set of positional moves where we could clearly say that move "X" is right... any other move is positionally worse. Or move X is wrong, any other move is better. But we don't have many of those. Since that is the case, I have always relied on my own chess judgement, plus (on occasion) the judgement of much stronger players, when I am not sure. I haven't found it hard to ask a GM "Why is this good or why is that bad?" And in general, I have not found their explanations hard to understand, although the occasional "it feels right" is a bit tough to program. :) The tuning approach I use is to simply play lots of games, extract the losses, and look for a "trend". If I find one, then I try to adjust things to repair what seems to be wrong...
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.