Author: Stuart Cracraft
Date: 18:26:23 12/28/05
Go up one level in this thread
On December 28, 2005 at 20:51:22, Dann Corbit wrote: >On December 28, 2005 at 19:56:16, David Dahlem wrote: > >><snip> >> >>>The idea works like this: >>>Suppose that I evaluate a pawn as 100.0 centipawns. How much should a knight be >>>worth? Probably somewhere between 200 and 500 centipawns. Now, if I choose a >>>bad value (like 200 centipawns) then I should solve less problems. If I choose >>>a good value (probably fairly close to 300), then I will solve more problems. >>>So I iterate over 12,000 quiet chess positions with known solutions and see what >>>value of the parameter gives me the most solutions. >>> >>>Now, the value found will NOT be optimal for play. But it will be a good >>>starting point for experimentation. And I will also have a pretty good idea >>>about what sort of range will produce reasonable results. >>> >> >>Thanks Dann. I'm not really a fan of tuning with test suites. I just build >>several test versions with different bonus values, and run a lot of test games. >>It's time consuming, but i have more confidence in the results. :-) > >I agree that test suites are not good for game play. But they are useful to get >an idea of the general ballpark for the right value of a parameter. > True - it is cheaper and easier to graft onto a program test-suite capability. But, for real study, I advocate using things like Genetic Algorithms, Temporal Differences, and other methods for tuning evaluation function parameters and coming to best results. I do not like to hand-tune as you can tell. The methods above may take as much time to initially program but can give unattended results that are as good as hand-tune (with less guilt/ongoing maintenance.) Stuart
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.