Author: Tom Kerrigan
Date: 16:35:50 07/25/00
Go up one level in this thread
On July 25, 2000 at 18:58:45, Pete R. wrote: >On July 25, 2000 at 14:12:46, Tom Kerrigan wrote: > >>It would be nice to make a change to your evaluation function and get immediate, >>accurate feedback. >> >>So my idea is to get a huge collection of positions of known value (i.e., "white >>has a stronger position") and then see how well the known values correlate to >>the evaluation function's scores. >> >>Does anybody have any ideas for getting a high-quality collection of such >>positions? Or any comments on this approach in general? >> >>-Tom > >I've thought about something similar, i.e. what if God handed you the perfect >eval for a number of positions, that would be wonderful. Obviously if you could >have an eval function that produced the same score that GMs would give a >position each and every time, your program would be unbeatable. But it would be >a fairly intensive project to get reasonably accurate evals on positions. How >many positions would you need? If it's a reasonably small number you might take >them from something with computer-checked GM analysis, such as Nunn's Chess >Openings. You could then cross-check those positions with a number of top >programs run to high depths to make sure no hidden tactics were missed. > >However, couldn't you simply use existing test suites for the same purpose with >just a little tweaking? For example suppose I am reasonably certain that a I've tried something like that. I wrote a program to go through a few dozen of Kasparov's wins. It did a 2-ply search on every position and gave itself a point if it correctly predicted Kasparov's move. At first it seemed like the number of points was useful and it was easy to tune my evaluation function to maximize the points. But then I found out that while I was maximizing the points for 2-ply searches, the number for 4-ply searches was dropping. So it was actually counterproductive. :( -Tom
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.