Author: Vincent Diepeveen
Date: 17:06:38 12/04/99
Go up one level in this thread
On December 04, 1999 at 01:35:52, Stephen A. Boak wrote: >I do not understand about assymmetric evaluations or giving a bonus for the side >to move. Please provide a brief explanation, thanks! > >>On December 03, 1999 at 21:31:09, Vincent Diepeveen wrote: >> >>Except for an experiment giving the side to move a bonus to some extend >>i've never been assymmetric. When I threw out this bonus too, then >>diep improved a lot in level. >> >>Logical one would say a bonus for having the side to move is good, >>but it never worked for DIEP. > >Why is it logical to give a bonus for having the side to move? > >1) If a position is zugzwang, then the side to move will lose regardless of >having the right to move. I detect zugzwangs in my nullmove already, remember? So those get detected already. We don't talk about a death&life bonus for having the move here. It's just another bonus out of the couple of the thousand that normally apply to an average middlegame position. >2) Having the right to move gives the side to move a choice of direction. >Having a choice of move allows the side the move to select its own destiny >(meaning path) but not necessarily its own ultimate destiny (meaning draw or >win). Having the move in general is a big advantage as you can do a move that gives you a position that is getting evaluated a lot better. > For example, if the position is bad for the side to move, how does having the >move in a bad position make it more likely that the position will improve >(meaning get better that it is at the moment) because of the move choice by the >side to move? > In a bad position, selecting the 'least bad' move among many bad moves does >not automatically make the position a good one. It simply minimizes the >negative aspects of the bad position after the move is made. This is of course true but making a move doesn't imply your opponent can already make a move. Quiescencesearch will see after making a move whether a too bad to be true. I have a lot of eval taking care of this type of positions especially. Positions where prospects are bad are statically evaluated a lot worse already than the bonus of the move. > Giving a bonus for the side to move might be overly optimistic for the side >with the bad position and the right to move. >Do I understand correctly--for each static evaluation of a single position (leaf >node, I guess), the program evaluates based on material and positional factors >without regard to which side is on the move? Then, some programmers like to add >a bonus for the side to move? It's an heuristic that's true in 99% of the positions that my program searches obviously. If a position is bad then usual programs nullmove. If a program nullmoves then the other side has the move so gets the bonus. So the sword cuts on 2 sides. Where you can nullmove normally to prevent getting into a worse position, you now suddenly cannot as your opponent gets the bonus instead of you. >> >>Never figured out why. Did i have a bug? >> >>Apart from this discussion, from which i don't know whether it's good to have, >>being assymetric becasue white might be a human and black a computer, >>i am against using an assymmetric evaluation function for that. > >Seems to me there are various types of assymetry in evaluation that a programmer >might try to incorporate (are these the kinds of assymetry you are discussing?): >1) The side to move gets a bonus (or penalty). Right. i'm just talking about this bonus. the other things i find objectively crap. other bonuses listed below are basically to cover some heavy weaknesses in evaluations of programs, or in case of evaluating moves are causing unbalanced trees. >2) The side to move, if in the better position, scores moves in a manner that >gives more contempt for a draw (goal is more to win, and less to draw--i.e., >more contempt for a drawing move/line). The side to move, if in the worse >position, scores moves in a manner that gives less contempt for a draw (goal is >more to seek a draw, and less to try to take risks to win). >3) Moves may be scored differently, depending on the perceived nature/strength >of a known opponent. Examples of differences in known opponents: human vs. >computer; human, maybe a GM, with certain propensity for playing some types of >positions well, and some not so well; computer with certain propensity for >playing some types of positions well--perhaps open; and some not so >well--perhaps closed or gambits requiring long term contempt for material in the >all-out seeking of a victory). > >Any enlightenment on the above subjects would be much appreciated. Then I can >follow the discussions with improved understanding and even more interest. >Thank you. > >--Steve Boak
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.