Author: Graham Laight
Date: 05:56:43 01/13/99
Go up one level in this thread
On January 12, 1999 at 12:57:18, José de Jesús García Ruvalcaba wrote: >On January 11, 1999 at 20:58:11, Graham Laight wrote: > >>On January 11, 1999 at 13:57:31, José de Jesús García Ruvalcaba wrote: >> >>>On January 09, 1999 at 05:55:25, Graham Laight wrote: >>> >>>>As I was sitting eating my breakfast just now, it occured to me that there are >>>>basically 3 items that, between them, will influence how close an evaluation of >>>>a chess position is to how good that position really is: >>>> >>>>1. The number of pieces of knowledge the evaluation function can call upon >>>> >>>>2. The quality of those pieces of knowledge >>>> >>>>3. The accuracy of selecting the right pieces of knowledge (and their >>>>appropriate weightings) for the position at hand >>>> >>>> >>>>Does anybody have any thoughts about this? >>> >>>I think that different evaluation functions are not comparable by themselves. >> >>Why not? >> >>You take a chess position, and run 2 different evaluation functions against it. >> >>The one that more accurately scores the position is the better evaluation >>function. >> > >Now the problem is, how to measure this accuracy? >There are only three posible theoretical values for a chess position (white >wins, draw or black wins), and it is unknown for most positions. An evaluation >function would be theoretically accurate if it gives every white win a better >score than any draw and every draw a better score than a black win, but I can >not imagine a way to find out other than solving the game of chess. > >Also, for a moment let us assume that the contempt factor is zero. If you take >an evaluation function and multiply it for *any* positive number, you get >different evaluation function which will *always* lead to the same best move! >Which one is more accurate? I agree with the sentiment - it is difficult to get perfect evals for most positions. This always seems to be a problem when I try to build expert systems - you don't know what knowledge will be most useful until you have built the system - but you can't build the system without knowing what knowledge will be most useful. >>>Overall program strength is. I mean, you can compare two evaluation functions >>>once you have all the other components of the programs fixed; but with a >>>different set of other components you can get different results. >>>Among the "other components" I can see: >>>1. Hardware: processor speed, and amount of memory used for hash tables. >>>2. The search algorithm, including extensions. >>>3. The opening book. >>>4. Endgame tablebases. >>>5. The time control. >> >>This is like saying, "You cannot evaluate the engine in a car unless you take >>into consideration the door handles and the headlights". >> >>I wanted to discuss the evaluation function of a program on its own - not the >>other stuff - important though I agree it is. >> > >Your original statement is essentially correct. I did not mean to disagree (in >fact I agree). My point is that I can not see a way to measure the quality of an >evaluation function by itself; but it is clear for me how to measure overall >program strength. > >José. > >>Ah well - I have to admit that sometimes it's the door handles that sell the >>car. >> >>Graham. >> >>> I think that the correct "accuracy" of the weightings can dramatically change >>>with these factors.
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.