Author: Robert Hyatt
Date: 20:46:44 01/04/98
Go up one level in this thread
On January 04, 1998 at 22:45:07, Don Dailey wrote: >On January 04, 1998 at 20:45:00, Robert Hyatt wrote: > >>I remain convinced that evaluation and search depth go hand-in-hand. I >>got burned badly once by not paying attention to this issue, when I >>really >>already knew about it. The story goes like this: > > >... > > >>The moral? The pawn hole code definitely improved CB at shallow depths. >> But >>CB understood outposts and weak pawns and so forth, and going from 4-5 >>plies >>to 9-10 plies made that pawn-hold code simply a "double penalty" because >>it >>got a penalty for the hole, plus it got a penalty when it saw how to >>utilize >>that hole. I've been careful since, although on ICC this is a problem >>since >>we play game in one minute as well as game in one hour or more... :) > > >This is a nice anecdote but I wouldn't take it too far. It sounds like >you're a strong believer in tunning your evaluation per level. Do you >think it will have to change significantly again when you are doing >15 or 20 ply? > yes. and there's another issue to boot. If your eval has 25 different features it evaluates, and the weights are all identical, at shallow depths it will probably play ok, because it has to choose only one of the features since the depth is so shallow. But if you search deeply, you find your code having to differentiate between (a) two of X and 3 of Y, or 1 X, 1 Y, 3 A's and a B, and so forth. So at deeper depths, your terms have to be more in line with reality, while at shallow depths just the mere presence of recognizing some feature is good enough... >I find your conclusions hard to believe although the story rings quite >true, I've had similar experiences. It's wise to suspect the last >changes you made when there is a problem. But saying this was good >code for 5 ply searches but bad for 10 ply searches is quite a stretch >for me, especially since you believe it was a significant amount in >both directions. I'm just guessing but is it possible you didn't >evaluate the initial code change properly? Now THAT would be easy for >me to believe because it sounds like someone we can all relate too! this was actual observation, not speculation, of course. IE we played *many* games (using the VAX) to be sure we liked the way it handled pawns, *after* adding the new code. And I'm talking about tournament time controls, with us watching the display as CB (Vax Blitz) searched. And then, after watching it play badly for one tournament and one game in the WCCC, we ran tests for almost 12 hours to convince us that CB with the code was making much worse decisions that CB without the 4 lines of code... We had no automated way of testing back then, because we could hardly ever get Cray time to test with... > >I have a similar story with bishop mobility. It looked like a great >thing for shallow searches and bad for deep searches. I was mislead >though by looking at the games. I see a position where the bishop >does a stupid thing and I'm ready to scrap it. It turns out the >algorithm hurt that position. But every heuristic will hurt some >positions because no one has perfectly balanced evaluation. They >are all guesses, seat of the pants judgements and never take enough >things into consideration. After really properly testing the bishop >mobility thing it turned out to always be good, it was just unclear >whether it was worth the extra time when you looked deep enough to >resolve the matter in other ways! > >But I don't think anyone really knows. It's very much a black art as >someone once said. That's why no two programs are the same. > >>I remain convinced that evaluation and search depth go hand-in-hand. > >I remain convinced that I don't know one way or the other. I think >finding out (and more importantly finding out WHY) will help all our >programs get better. > >Do you have some thoughts on which evaluation weights should increase >or decrease with depth? If you really believe this stuff, help me >come up with a body of theory that describes and predicts it. I'll >offer a starting proposal which you guys can refute or improve on: > > 1. Material > 2. Pawn structure > > 3. King safety > 4. Time > 5. Space > >Some masters break chess down into these 5 elements. Material and >Pawn structure are considered Static features and the rest are >considered >dynamic. Static features tend to be more long lasting and do not change >too much with equal material exchanges, but dynamic things change >quickly >with equal exchanges. Another way of looking at them is that you may >notice you are always trying to convert 3,4 and 5 into 1 and 2. King >safety you try to convert to checkmate which is the ultimate win of >material, but often your opponent will avoid getting mated by giving >you some material. > >My hypothesis is that your program should de-emphasize >dynamic features with increasing depth and emphasize static ones. We >will >consider trying to figure out pins and forks as "TIME" which is a >dynamic >feature. Time is initiative and involves threats and such. My own >experience is that resolving these does seem to be worth more at shallow >depths but worthless with deep searches (despite my skepticisms.) > >The 5 elements can be broken down many ways and some masters use 4 >elements >not counting king safety and I'm sure we could break them down finer if >we wanted too. But this is the general idea. > >What do you think? Pure bullshit or is there something to this? > > >-- Don I've long felt that pawn structure is *the* important feature. Because something broken often can't be fixed... while a cramped piece can be remedied over time. So you might be right here too, that things like "get the pieces closer to the king" are not as important at deep depths (because if you search deep enough, you see *why* they should be closer) while pawn structure changes will be beyond any search for a long time...
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.