Author: Robert Hyatt
Date: 10:12:46 01/20/06
Go up one level in this thread
On January 20, 2006 at 10:22:10, Michael Neish wrote:
>On January 20, 2006 at 09:44:43, Steffen Jakob wrote:
>
>
>>If you want to experiment with this idea you can reduce the value space of
>>your evalation function by something like this:
>>
>>Value Engine::FilterValue(Value v) const {
>> return v - (v % M_eval_filter);
>>}
>>
>>First you evaluate the position and after this you filter the computed score.
>>E.g. if M_eval_filter == 5 you only get scores which can be divided by 5.
>>
>>Best wishes,
>>Steffen.
>
>Interesting suggestion, thanks. It certainly does reduce the tree size when
>using a value of 3, but not 2. Whether and how it impacts the quality of play
>is another matter. It's certainly different enough to make the program choose a
>different first move (without book).
>
>Cheers,
>
>Mike.
What you are doing is collapsing all scores within a small window to a single
value. As I mentioned in the previous post, the less variance in your score,
the smaller the tree. But then you run into random variation, because since we
prune on <= or >= tests, the _first_ move that produces one of those discrete
scores is going to be best and all others will fail the alpha/beta test. That's
why material-only will play some odd moves, most notably the move sorted first
at the root will almost always be played unless it loses material and gets
replaced by a move with a better root score. But if all root scores are zero,
the first is best according to alpha/beta
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.