🚀 go-pugleaf

Tom Vijlbrief (tom@izf.tno.nl) wrote:
: I noticed when running some of the Bratco-Kopec tests that
: GNUCHESS shows some strange behaviour.

: The idea is to turn off the hashtable and let it play a move:

----- much text deleted ------

: =====================================
: Note that at the first move computation ply 2 shows a tree count of 4431 .
: After the undo the tree count at ply 2 is 4437 !!!

: Does somebody has an explanation for this. I think that somewhere
: a variable is uninitialised !!!

: This is more important than it seems at first sight
: because after the undo, gnuchess sometimes plays a much better move !!!


The difference is due to the static cache which is not cleared following
the undo.  If you were to add a ZeroTTable () in SelectMove () before the
iterative deepening begins, you will find that the inconsistency will 
disappear.  (I've tested this of course).


: I discovered this while implementing a better null move heuristic
: and some search extensions. These look very promising but before I
: release the diffs I would like to solve this problem....


Now that the problem is solved, you can release those diffs.  It would of 
course be better if you were to send them to Mike McGann at 
mwm@hasler.ascom.ch so that he can incorporate them into pl 63.


Kong Sian
kongsian@nsrc.nus.sg