Author: Robert Hyatt
Date: 10:48:51 06/25/98
Go up one level in this thread
On June 25, 1998 at 10:37:12, Robert Hyatt wrote: >On June 25, 1998 at 09:30:58, Ed Schröder wrote: > >>>Posted by Amir Ban on June 25, 1998 at 03:36:32: >> >>>I use the same procedure Don uses: Two killers, always replace, no counters. I >>>was surprised you guys think you have something better. Tests will decide, >>>true, but I'm not persuaded by the verbal arguments. It seems to me that >"always replace" should be equal at least. >> >>Same here. Tried all types of counters and other tricks to no avail. Always >>replace for Rebel is also superior. >> >>Like to add that the following worked for Rebel. After killer-one I use >>killer-one of 2 plies back, then killer-two and finally killer-two of >>2 plies back. This gave a speedup of 5% if I remember well. >> >>- Ed - >> >>>Amir > > >that worked in Cray Blitz, but we didn't use history moves there. In Crafty, >I try the two killers, then a couple of history moves, and if I haven't gotten >a cutoff by then, I stop wasting time and just take moves as they were >generated, to save time, since this is probably an ALL node anyway and won't get >a cutoff. > >My only comment on the counters is that it works better for me. It did in >Cray Blitz, and it does in Crafty. I tried both ways way back, but this was >always a little better... I'm not sure what is meant by "extra overhead of >counters" because it adds basically nothing. You don't have to "sort" with >two entries.. you always replace the last, and when you get a match on the >second one only, you bump the counter and possibly swap it with the first one >based on the two counter values. Not exactly time consuming. I just ran some tests, since I haven't done this in quite a while. My results haven't changed... some positions are faster with the count, others are not, but overall, the counter approach produces results that are between .5 and 1% better. I notice *no* speed penalty with the counter, and my NPS didn't change at all, but the trees *overall* were a tiny fraction larger. Note that I tested only a couple of tactical positions, most were just common middlegame positions, I suspect that this is one of those things where *any* sort of killer idea is better than *no* killers at all... This is probably along the lines of hash probing.. do you use a two-level (Belle) approach (one always store table, one depth-preferred table) or do you probe multiple positions? Multiple probes are clearly better, but they have a cost. If you can afford huge, sparsely-populated tables, the two-level approach can work well, because it strains the memory bandwidth badly. The cheaper, and less-bandwidth-intensive two-level table can do pretty well if it is not too small. And the approach (whatever it is) is a near-religion to many. :) Bob
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.