Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Building the Principal Variation in MTD(f) searches

Author: Dan Andersson

Date: 11:27:58 07/18/99

Go up one level in this thread


>We all don't doubt that MTD is a fine algorithm. Regrettably it only
>works for programs using an evaluation which gives very little deviation;
>
>See the score differences the first few ply in DIEP's lines:
>
>0.87
>1.21
>0.63
>0.87
>1.29
>0.93
>0.94
>0.94
>1.00
>1.03
>0.77
>0.78
>0.82
>0.89
>1.03
>0.80
>0.81
>0.80
>0.86
>0.84
>1.03
>0.65
>0.97
>1.04
>1.20
>1.02
>1.15
>0.88
>0.98
>0.86
>
>And those deviations still considering it's a position where there is
>not too much tactics.
>
>Getting the big fail low at 14 ply caused a jump from 1.03 to 0.65
>which took 28.00M / 12.49 = 2.24 branching factor
>
>Just imagine the huge number of researches that MTD needs
>to get to 0.65!

Probably one or a few re-searches only, I have not noticed that drawback in my
project. The problem is to implement lazy evaluation, thus the correlation
between pawn-values and evaluation as you mentioned.
I use a alpha-betaized evaluation function (it orders the evaluation factors in
a binary search and cuts off when it is out of alpha or beta boundaries and
returns alpha or beta) and save the state of the evaluation function in the hash
table if the position was not fully evaluated, to continue evaluating later if
needed. This way I get the benefits of MTD and lazy evaluation. i.e alpa-beta
evaluation cut evaluation time by 75% and it allows bootsrapping evaluation
values.

>
>Because let's face it, what program *actually* use MTD currently:
>
>parallel version of fritz (?) :  evaluation of it under 200 clocks
>cilkchess                     :  evaluation also is very limited and
>                                 definitely doesn't have huge scores,
>                                 we all know that cilkchess evaluation is
>                                 always very close to the number of pawns that
>                                 it sees.
>SOS                           :  same comment valid as for cilkchess
>
>So the programs using MTD currently have hardly anything in their evaluation
>not many things causing big scores,
>so claiming that MTD is doing a good job is only true when taking into
>account this restriction, which a lot of programs do not wish to have.
>
>Greetings,
>Vincent



This page took 0.01 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.