Author: Steve Maughan
Date: 06:13:17 12/24/02
Go up one level in this thread
Dieter, >With your formula, a problem might be, that it has a time domain. Yes it is and I see that as a benefit! If you set the half life to a low value and steadily increase it you will see the total score, expressed as a percentage, increase. This reflects that given more time, programs do play stronger moves. >Assume I >change a version, and test again. All will be the same, but one solution time >goes from 30 seconds to 60 seconds. This will lose about 2 points. Another >solution goes from 480 to 180 seconds. This will gain about 1.2 points. My >feeling would say, the later version is not worse, but with the given formula, >it is. When just comparing the sum of the logarithms and the number of solutions >stays the same, to me, the comparision would look more natural. In the example >above, the later version would win. I see however, that your formula also has >advantages. It handles the cases of different number of solutions better. To me >it looks, like it would work best, when the typical solution times are not too >far from the half-live time. I have also applied this type of formula to positional test suites. As an example suppose that a position had a list of possible moves with some form of positional score associated with them e.g. Ng5 = 10 points Bxe2 = 7 points Qh2 = 6 points Qh3 = 6 points g4 = 4 points Nd2 = 1 points Suppose that a program considered g4 to be the best for the first 45 seconds, then switched to Qh3 and finally switched to Bxe2 after five minutes. The score for this position would be calculated as: Score 1 (g4) = 4 * (exp(0 *ln(0.5) / HL) - exp(45 * ln(0.5) / HL) Score 2 (Qh3) = 6 * (exp(45 *ln(0.5) / HL) - exp(300* ln(0.5) / HL) Score 3 (Bxe2)= 7 * (exp(300*ln(0.5) / HL) Position Score = Score 1 + Score 2 + Score 3; It works well - all you need is a set of good quality positions and assocaited positional moves. Regards, Steve
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.