Author: martin fierz
Date: 02:53:21 07/02/03
Go up one level in this thread
On July 01, 2003 at 14:10:44, Robert Hyatt wrote: >On July 01, 2003 at 05:41:31, Graham Laight wrote: > >>When a GM is contemplating a move, he doesn't say to himself, "Hmmmmm. I would >>give the resulting position a score of 1.723". > >Actually, he _does_. > >"Nb5 drops a pawn but has compensation in the attack on the enemy king." > >"Nf3 maintains material equality but my position is quite cramped." > >etc. > >I think that a human does _far_ more than just say "this is equal, winning >or losing." I know I do. And the GM players I talk to seem to do this as >well. While they might not do millipawns (or even centipawns) they certainly >do fractions of a pawn in positional compensation. i think the main difference in human/computer eval is not what the original author thinks. i agree with bob that players do think in such terms, e.g. i gave up a pawn and i have "about enough compensation" or "some compensation but too little" or "more than enough compensation" for it. this is more or less a direct translation from numbers to words or vice versa. the real difference lies elsewhere. humans can look at a position and say things like "this is probably better for white but it's very unclear" or "this is a technical position which is either a white win or a draw". when a human has a choice between the above two possibilities, it is a matter of style which one he chooses. a computer might give both of these positions the same number, and does not choose based on style. cheers martin >>Such an evaluation is nonsense anyway. There should properly be only 3 >>evaluations: >> >>1. Winning position >> >>2. Drawing position >> >>3. Losing position >> > >That would be great if it were possible, but except for forced mates and >forced repetitions, there are no such "exact evaluations" in the actual >game. A GM might say "this is winning" but it is based on very fuzzy >"computation" done mentally based on past experience and preferences. > > > > > >>It would be nice if a program could work as follows: >> >>"nb5. This position contains a possible bishop trap". >> >>"nd5. This puts more pressure on the opponent's king" >> >>"Opponent classification: bishop trap success rate = 25%" >> >>"Opponent classification: king attack success rate = 15%" >> >>"Choice = nb5". > >That is about "discernability". It is a tough problem but a well-known >issue in computer chess. > > > >> >>-g
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.