Hash table use in Chess Quiescence Search
Author
I''m looking for ways to speed up my quiescence seach function (refered to from here on as qsearch).
Generally my function works as follows:
a. Try a static eval of the move-- if a cutoff occures, return the static eval.
b. Otherwise, explore all captures.
My current version of the qsearch does not make use of the transposition table and it seems like this would be the best place to start.
There are two issues here that I''m not quite comfortable with:
1. The qsearch is not a full-width search.
2. The qsearch does not have a reliable indication of ''depth'' searched. It could cut off early, or not, while the main Alpha Beta search is a definite ''ply 7'' or ''ply 6'' search.
I''m affraid that simply storing qsearch values in the transposition table will screw up the main Alpha-Beta search. Ideally I''d like to be able to exit from the qsearch early based on move transposition.
I''ve read comments by others suggesting that there is no danger in treating the qseach just like the main Alpha Beta but my gut tells me I''ll run in to a bad case of search-instability if I do this.
Are there theoretically sound forward pruning techniques that can be used in a quiescence search?
Thanks,
Will
If all AlphaBeta end nodes are evaluated with qsearch() then I don''t see the problem. I am however a bit concerned about your performing a regular eval() first and trying to do a cutoff.
The point of the qsearch() is to avoid horizon effects that eval() simply can''t handle.
The point of the qsearch() is to avoid horizon effects that eval() simply can''t handle.
quote:Original post by Anonymous Poster
If all AlphaBeta end nodes are evaluated with qsearch() then I don''t see the problem. I am however a bit concerned about your performing a regular eval() first and trying to do a cutoff.
Performing a regular eval() first is absolutely needed for a quiescence search to work. A side should not be penalized because all the captures available are bad. Think about it.
As for the original question:
1. The qsearch not being a full-width search should not be a problem. Hash tables can store information on bounds, aswell as exact scores.
2. You can use a depth of 0 for nodes in the quiescence search. This should not create any more search instabilities than any other hashing.
To make your quiescence search have as few nodes as possible, make sure that you sort the moves correctly (capture major pieces first, with smaller pieces first). You can also use futulity pruning, although it has never worked for me (probably because my evaluation function is too wild).
Author
Alvaro: Thanks for clearing that up-- in my own program I do the static eval first, check for a cutoff, and then try captures. I'm using MVV/LVA to sort the moves.
So every node in the qsearch is stored as a depth 0 search? I should have just tried that to begin with.
Thanks,
Will
[Alvaro had posted while I was typing this the first time]
[edited by - RPGeezus on June 7, 2004 9:21:18 PM]
So every node in the qsearch is stored as a depth 0 search? I should have just tried that to begin with.
Thanks,
Will
[Alvaro had posted while I was typing this the first time]
[edited by - RPGeezus on June 7, 2004 9:21:18 PM]
Another tip to speedup qsearch, especially if you have a non trivial eval.
You can use an "eval cache" instead of storing depth zero in the hashtable. Think of it as a specialized hashtable that store only evals. You probe it in your eval() function and store the result there was well. Depending on how your search look and other heuristics behave (the main hashtable effeciency mainly), in my experience, you can have from 30% to 50% of probe success, that means 30%-50% less evals to do. Plus you don''t "pollute" your main hashtable with 0 depths results. Of course it has drawbacks (more tables access) so it''s all about testing and testing. Try it and it doesn''t work dump it.
You can use an "eval cache" instead of storing depth zero in the hashtable. Think of it as a specialized hashtable that store only evals. You probe it in your eval() function and store the result there was well. Depending on how your search look and other heuristics behave (the main hashtable effeciency mainly), in my experience, you can have from 30% to 50% of probe success, that means 30%-50% less evals to do. Plus you don''t "pollute" your main hashtable with 0 depths results. Of course it has drawbacks (more tables access) so it''s all about testing and testing. Try it and it doesn''t work dump it.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement