Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Evaluation-based Reductions and/or Extensions

Author: Tord Romstad

Date: 05:44:15 12/29/03

Go up one level in this thread


On December 29, 2003 at 03:47:11, Vasik Rajlich wrote:

>On December 28, 2003 at 21:07:34, Tom Likens wrote:
>
>>On December 28, 2003 at 16:55:49, Tord Romstad wrote:
>>
>>>Hi Tom!
>>>
>>>That's precisely what I am doing in Gothmog, as you probably know.  I'm glad
>>>to hear that you try similar ideas, and that your first results are promising.
>>>I've found that it takes lots of tuning and testing before it works well,
>>>but I think such techniques have great potential in the long run.  If it's
>>>done right, most of the improvements you do in your evaluation function will
>>>automatically improve the accuracy of your search.
>>
>>When I first started experimenting with this idea I was *very* aggressive.
>>Interesting it didn't really hurt the programs tactics and it seemed to
>>reach much deeper depths while searching, so I was very excited.
>>Unfortunately, when I played the new version against the older non-pruning
>>version it lost rather badly.  So, now I'm starting out in a more
>>conservative fashion and tuning the reductions based on actual games.
>>
>>>>My (obvious) question, how do other programmers deal with this phenomenon?
>>>>I suppose ignoring it is one option, but I'm hoping there is a better
>>>>solution.
>>>
>>>I agree that this problem is extremely annoying, and I have spent lots of
>>>time and effort trying to find a solution.  Unfortunately, I still haven't
>>>found any good ideas. I asked a question about exactly this problem here in
>>>CCC just a couple of days ago, but the only person who replied was Dieter
>>>Bürssner, who also hadn't found a better solution than just ignoring the
>>>problem and hoping it wasn't too important.
>>>
>>>Tord
>>
>>The fact that neither yourself nor Dieter have found a satisfactory solution
>>illustrates the difficulty of the problem.  It may be possible to tag these
>>nodes when they are saved into the hash table and simply use them for move
>>ordering, as Uri suggested.  I need to gather some data before I can make
>>any kind of intelligent decision.  I do agree though that the concept has
>>*massive* potential.  I wouldn't be surprised if this wasn't part of the
>>"secret" of the commercial programs, (especially that Tiger fella) ;-)
>>
>>--tom
>
>Hello,
>
>It seems to me that there is a way to deal with this.
>
>Let's say you're at node A going to node B. You can statically evaluate node A,
>statically evaluate the move (A->B) in the context of node A, apply the
>appropriate reduction/extension, and enter node B with the already
>extended/reduced search depth. In this case, node B will go into the hash table
>with that new remaining search depth.

As far as I can see, this doesn't solve the problem, but just moves it to a
different
node.  Consider the simplest and most well-known path dependent extension:
The recapture extension.  The move (A->B) could be a recapture in one node and
not be a recapture at a different node with the same position on the board.  The
node
B will be stored correctly in the hash table, but there is a problem with node
A.

Tord



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.