🚀 go-pugleaf

Dan Oetting wrote:
> 
> The whole idea of move sorting is to find the best move sooner and
> therefore reduce the node count with alpha-beta cuts. You cannot measure
> the effects of move sorting by looking at nps, you need to look at time
> to ply.

Quite right - although if I cut the history code it does examine nodes
quicker as well as solve more of my test cases. So either the history
code is examining the wrong nodes first or it is not worth the CPU time
spent on it - suspect it is either too crude, or just not useful at
blitz speeds.

For the main test case it actually looked at less nodes without this
term in the sort, and solved more puzzles at the same time control.

                30/62 13703081 A hacked about gnu 5.3 before
                32/62 13540009 and after cutting history.

This is based on how many nodes examined, and how many puzzles solve on
the last 62 studies in Larsen's 63 end game studies for Grandmaster at
5s/move.  

Not sure 2 games in 62 is significant - but this set of tests has proved
pretty good so far.   

I've sent Stuart a patch to make it try the first position in an EPD
file, that was the easiest way to improve it's score by one 8-)

	Simon
-- 
Business http://www.eighth-layer.com/
Personal http://www.wretched.demon.co.uk/