Author: Dann Corbit
Date: 11:22:43 04/05/04
Go up one level in this thread
On April 05, 2004 at 08:26:08, Andrew Wagner wrote: > >I'm taking this intro to calc-based statistics class this semester, and it got >me thinking. We tend to generate a lot of statistics during a game. In fact, we >could generate a LOT more. But we do very simple things with these statistics. >For example, we might add some bell or whistle to the search, notice that it >doesn't show any obvious increase in NPS or move ordering, or whatever we were >hoping for, and ditch it. That's an extremely simple analysis of the statistics. >I'm wondering if anybody has done any more advanced statistical work in figuring >out what factors are important to winning a game. Here are two examples I >thought of: > > 1.) Using advanced statistics to calculate better values to be used in >static eval. What if we took a ton of super-gm games, and generated a slew of >statistics about each game like number of moves white has an isolated pawn on, >or number of moves black keeps his king-knight, and so on. Then do a >multi-variate analysis to determine whether any of those things have an effect >on the probability that white or black will win. Unfortunately, this math is a >little above me at this point, and maybe someone else has tried it, I don't >know. It's an interesting idea to think about, though. > > 2.) Some quantitative analysis of the speed vs. knowledge question. We have >the speed number readily available, in the form of NPS. If we could find some >way to quantify the amount of knowledge we build into an engine, we could easily >find some kind of ideal balance. Think about it: this kind of thing is done all >the time in many fields. Mathematically, this kind of optimization should not be >all that hard, if we can come up with good numbers to use. > > So, what do you think? Anybody heard of such a project before? Do you all >think it would work? I'd be interested in any thoughts you have. Andrew I have code that does a parabolic least squares fit for evaluation parameters. You can tune it to do anything that way, but the problem is --> that will be what it is good at. So (for instance) I can tune it to test suites and it will slaughter them, but in real games it takes a pounding.
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.