Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: MTD(f)

Author: Vasik Rajlich

Date: 11:01:36 06/18/05

Go up one level in this thread


On June 18, 2005 at 06:51:45, rasjid chan wrote:

>
>QUOTE-
>Actually, it's really not that complex. (Unless I am missing something really
>stupid :))
>
>You don't need any boundary information, the fact that the score is outside of
>the alpha-beta bounds implies where you are in relation to the boundaries. The
>difference between fail-soft and fail-hard is something like 4 or 5 lines of
>code. You're just not restricting your return values to alpha-beta.
>UNQUOTE.
>
>Some time back, you gave me some tips on fail soft and after that
>I seemed to know how to do it correctly and I have the idea
>that what I got is "correct". I have to really think on your "4 or 5 lines of
>codes" and if it is true that that is also possible and yet as complete as
>what I feel I get with my implementation, I would be surprised. Never mind about
>explaining the details, I'll see if I am doing something silly, then I may post
>for help. Maybe others may just give a short comment if actually
>fail-soft may be done with some clever short codes and yet enough.
>

Probably you also did it correctly - there should be several solutions.

>About what I monitored, I just count the number of fail-low and average the
>deviation from alpha. It is about 1 pawn for the whole game.
>

Aha. So you combined the results from searches of all depths. If you separate
your searches by depth, you will find that low-depth searches tend to return
softer values, while high-depth searches tend to find a way to return values
closer to the bounds. For a number of reasons, search has a certain natural
laziness :)

Vas

>Of course I implement fail-soft just knowing it is better, but not knowing
>how important. Just like in the thread about why is fruit strong,as you
>mentioned, an overall balance and proper priorities.
>
>Best Regards
>Rasjid



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.