Author: Ricardo Gibert
Date: 10:37:32 11/19/00
Go up one level in this thread
On November 18, 2000 at 17:57:51, Daniel Kang wrote: >Hi there! > >On November 18, 2000 at 06:03:39, Graham Laight wrote: > >>IMHO, a truly accurate evaluation of a position would yield one of the following >>3 ordinal values: >> >>Win >>Draw >>Lose > >Wouldn't it be even better if it gave the expected value of a match? Since, in >practice, chess is a game of incomplete information (at this point, chess is >solvable only in theory), isn't it better to judge a position by the possible >outcomes and how likely each outcome is, rather than its absolute theoretical >value, which neither the program or its opponent can be expected to compute? Consider K+R vs K. Most legal moves by the superior side are winning moves and the expectation of winning for each of those moves are about 100%. How do you tell the difference between the moves that bring you closer to mate and the moves that are winning, but prolong the game? The term "evaluation" is something of a misnomer and as it used in computer chess, which is significantly different from how human players use it. A better name is perhaps "preference-score". Naturally, a program will/should have higher scores for those positions that bring the game closer to mate. This is what is "preferable". > >I've been thinking of this evaluation idea for a while. It seemed to me that >many strategic errors made by computer programs can be attributed to their >inability to compare positions that are wildly different. Obviously, being up by >two pawns is different in a wildly tactical warfare than it is in a quiet >endgame. It's been a while since I looked into the state-of-the-art of chess >programming, so I might be a bit off, but it doesn't seem to me that most >programs are anywhere near accurate in adjusting for these differences well. > >Dan.
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.