Author: Stuart Cracraft
Date: 16:03:53 09/26/04
Go up one level in this thread
On September 26, 2004 at 04:30:44, Ed Schröder wrote:
>On September 26, 2004 at 00:51:17, Sergei S. Markoff wrote:
>
>>Some ideas on hash move extensions.
>>The idea was invented independently by Shpeer, Rybinkin ("Mirage") and me. In
>>most cases this is a good idea to extend TT moves to improve search stability.
>>But what scheme will be best for doing it?
>>
>>In ST the method is a little curious.
>>
>>TT_extesion_amount=INCPLY/RootDepth (in plies).
>>while(TT_extesion_amount*RootDepth<INCPLY) TT_extesion_amount++;
>>
>>Anyway it improves play of ST.
>>
>>I need your results to compare.
>
>Since Rebel 12 I am using this:
>
> in case of (2 consecutive best_move_from_hash_table
> && alpha < hash_table_score + margin)
> then { extend with one ply in case this situation happens x times }
>
>Notes:
>
>. margin is currently set to 1.00
>. x is currently set to 4, 8, 12, 16 etc.
>. alpha in Pro Deo's philosophy could be beta in yours.
I tried this -- hopefully I did it. If 2 consecutive hash table
probes give a best move and beta < hash_table_score_returned + 1 pawn
then extend.
I assume your ply increment is in 1/4 hence the 4? so I am not sure
I did it right.
My extension is based on things like extend+=0.25, extend+=0.50, extend+=0.33
etc. so floating point. And mine simply does this:
If 2 consecutive hash table
probes give a best move and beta < hash_table_score_returned + 1 pawn
then extend to a max of 1 ply.
I tried this on my standard suite and saw no improvement but not something
good must be there -- what should I be looking for to confirm positive effect.
Thanx.
>
>The background is simple, in case you have 2 consecutive
>best_move_from_hash_table and the score is reasonable you most probably are in a
>part of the tree with move sequences that make sense, thus extend.
>
>One advantage of the algorithm is that not only the usual situations (checks,
>re-captures etc.) are part of the extension but also the "normal" moves.
>
>If you have Pro Deo then check the MISC_37 parameter and its sub-parameters for
>the effects of the algorithm, the implications are simply huge.
>
>[Pruning = MISC_37] * Hash table extension
>[M37_VAL = 1.00] * margin is 1.00
>[M37_count = 4] * in steps of 4
>
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.