🚀 go-pugleaf

Roger Uzun (uzun@crash.cts.com) wrote:

: On my machine KILLT is faster than HISTORY, and I thought they
: were equivalent logically, does anyone know if HISTORY
: is indeed preferable to KILLT, despite the speed penalty
: caused by using HISTORY (even with the fix/enhancement
: posted here by Kong Sian).

: -Roger
: --------------------------------------------------------------
: bix: ruzun
: NET: uzun@crash.cts.com

KILLT and HISTORY are not equivalent.  Both of these are basically move
ordering heuristics.  KILLT orders moves by placing killer moves (PV,
killr0, etc) first.  It also favours moves towards the centre of the board
(see Initialize_killt).  History keeps a score of the best moves in the
interior nodes based on the depth of the sub-tree search. If at an
interior node, mv is the best move obtain from searching this node to a
depth of d, then the score for this move is increased by 2**d.  Generated
moves are then ordered based on this score. 

I did a simple experiment, based on the suggestions that I've made
regarding KILLT and HISTORY.  The first 10 positions from the larsen1.fin
test suite is taken and searched to 5 plies.  Here is the result.


	USING KILLT & HISTORY			USING KILLT

1.		401560				415906
2.		3233				3911
3.		308222				334665
4.		65456				64858
5.		311220				380129
6.		62486				63861
7.		234292				233242
8.		119294				120006
9.		146577				145673
10.		3675				3457

Total:		1656015				1765708

The numbers refer to the total number of nodes searched.  As can be
seen, using history gives about 6% less nodes.

Kong Sian