Computer Chess Club Archives




Subject: Re: MTD: an observation and a question

Author: Tony Werten

Date: 00:40:47 09/13/02

Go up one level in this thread

On September 12, 2002 at 18:30:07, martin fierz wrote:

>On September 12, 2002 at 16:50:38, Dieter Buerssner wrote:
>>On September 12, 2002 at 16:10:38, martin fierz wrote:
>>>bestvalue = -mate;
>>>for all moves do
>>>  {
>>>  value = -search(-beta,-alpha,etc);
>>>  bestvalue = max(value, bestvalue);
>>   alpha = max(value, alpha);
>>>  if(value>beta)
>>>     break;
>>>  }
>>>return bestvalue;
>>>is there anything wrong with this?
>>Pehaps you left it out on purpose - but this code snipped is missing an update
>>of alpha. For example at the place, I suggested above.
>hi dieter,
>you are right, i left it out, half on purpose i guess. it just seemed irrelevant
>to the return value. but it's in my code - although it's really unnecessary: in
>MTD you always fail, so you never have to update alpha. but i have a compile
>switch to go from a windowed search to MTD, and i never bothered to write an
>MTD-only negamax function, although i could save an if or two there :-)
>>I see the same problem as you. Even when using strict fail soft search (I also
>>consider other points mentioned in this thread, like lazy eval), I typically
>>(but not allways) get back a score of X+1 for search, that fails high in the
>>window X, X+1. I think, your suggestions "alpha-beta tries to do the least work"
>>explains this. With a deep depth, and a complicated search tree, the X+1 score
>>will be easiest to prove and most probably be proven first.
>i just made some more tests in positions where the score really swings much
>(equivalent to winning a pawn in chess) and now i have seen jumps of more than
>the eval grain, although only 2 or 3* eval grain, like the (0.00,0.01) search
>returns 0.03. but most of the time it jumps as little as possible.

This seems quite normal. I sometimes manage to get checkmate scores back to the
root, but it doesn't happens very often. The deeper the searchtree below a move,
the more likely the checkmatescore is "lost" somewhere.


>  martin

This page took 0.05 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.