Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: MTD(f) and storing the PV

Author: Tord Romstad

Date: 13:44:28 07/29/03

Go up one level in this thread


On July 29, 2003 at 15:50:18, Andrew Williams wrote:

>On July 29, 2003 at 14:58:48, Tord Romstad wrote:
>>There is still a long way left before mtd(f) works as well as traditional
>>aspiration search in my program, but I will continue trying for some time.
>>mtd(f) is fun and addictive.  :-)
>>
>>Tord
>
>Yes. There's just so much scope for experimentation. For example, in Aske
>Plaat's online description, he raises and lowers the guess by 1 each time
>around the MTD loop. I've never been able to make this work effectively. Far
>better to accelerate this. If the score is going up, I add 1 the first time
>then 2, then 3 etc etc. I think I stop at 7 or something. I can't remember
>now. The same applies to when the score is sinking.

I was planning to do something similar, but so far I have only implemented
the algorithm almost exactly as in Plaat's online description.  I have taken
my old alpha beta search, converted it to fail-soft, removed all extensions,
reductions, pruning (including null-move) and all other tricks, and removed
checks from my qsearch.  I use the mtd(f) driver with this simple and
primitive search.  Of course it is very slow and doesn't play well, but it
is easier to debug the mtd stuff if I keep everything else as simple as
possible until it works.

My biggest problem right now is that for some reason the PV (even the
first move of the PV) is often wrong immediately after the program finds
some tactical trick, and that the right PV only appears at the next
iteration.  For instance, in WAC-1, which is a mate in two moves, my
reports "mate in two" at search depth 3.  Because the program has no
extensions or pruning whatsoever, this is precisely as expected.  However,
the right move appears only at iteration 4.  I have seen this phenomenon
in many different positions.

Even if I end up rejecting mtd(f) (which is not very likely), my experiments
have been worth the effort.  It is a great way to eliminate hash table bugs!

>Have you read Aske Plaat's PhD thesis? It's available online somewhere and is
>very interesting and readable.

I have found it and downloaded it, but haven't started reading it yet.
At this stage, his short online description is enough for me.

Once again, thank you very much for your advice!

Tord



This page took 0.03 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.