Author: Tony Werten
Date: 01:43:05 06/26/02
Go up one level in this thread
On June 25, 2002 at 20:14:39, Robert Hyatt wrote: >On June 25, 2002 at 17:42:58, Tony Werten wrote: > >>On June 25, 2002 at 17:30:51, Ulrich Tuerke wrote: >> >>>On June 25, 2002 at 02:40:59, Gian-Carlo Pascutto wrote: >>> >>>>On June 24, 2002 at 18:53:24, Steve Coladonato wrote: >>>> >>>>>>I wonder what you consider 'comparable'. There's no guarantee >>>>>>they'll be similar whatsoever. >>>>> >>>>>That was not a well formed statement on my part. What I meant was that for a >>>>>given ply depth, the evaluation that program X comes up with should be >>>>>comparable to the evaluation that program Y comes up with if both programs are >>>>>fairly equal in overall strength. >>>> >>>>No. There is no guarantee whatsoever that this is true. >>>> >>>>>Therefore, if the algorithms/heuristics that >>>>>program X uses allow it to get to ply M faster than program Y, then program X >>>>>should win if the time allowed constrains how much time each program can use for >>>>>analysis at that depth. For example, if program X can get to ply 11 in 30 secs >>>>>and program Y takes 1 min 30 secs to get there, the overall analysis that >>>>>program X can generate during a game should be better than that generated by >>>>>program Y and program X should win. So it seems that the efficiency of the >>>>>algorithms/heuristics will determine the overall strength of a program. >>>> >>>>Again, this is completely false. >>>> >>>>I will repeat what I said several times earlier in this thread, and that >>>>is that plies are not comparable between chessprograms. The analysis of >>>>one program at ply 11 can be completely different and of higher >>>>quality than another at the same 11 ply. If the second program reaches >>>>ply 11 faster, we have no information at all to make any solid conclusions >>>>about the relative strength of those programs. >>> >>>Completely agreed. This integer which we are talking about should be better >>>called "iteration number". It basically defines how many times the search had >>>been restarted exploiting each time the results of the preceeding iteration in >>>order to extend the search tree. >>>IMHO, the relation of iteration number to search depth is a very loose one, >>>having in mind that todays programs are heavily pruning as well as extending. >>> >> >>Hmm. I can imagine that a program that uses partial ply extensions might decide, >>when the timelimit is almost reached, to start an iteration with only half a ply >>deeper. >> >>Or even worse. Every uses iterative deepening, but did anybody ever prove that >>full plies are best ? Maybe 2/3 ply is better ? >> >>Tony > >I didn't "prove" it, but I did test a bunch of different increments a few >years ago, from .5 to 2, and liked 1 the best. Sometimes going by .5 would >go instantly since 1/2 ply is not really an extension unless something else >gets added. The bad thing was that on occasion, critical hash info would >cause the N+.5 search to take longer than necessary since the search had to >be re-done if the hash table was overwritten in a critical spot. Then that >would be wasted... > >I didn't do exhaustive testing however, just a bunch of positions with various >increments from .5 to 2.0... Everyone should try it to see if they find that >something other than 1.0 is better for their program. And once they do, they >probably should re-test yearly to make sure the answer is still the same then. > I think it could be nice for the "high" iterations. First because they tend to have a higher branching factor and second because there will be a lot of plies that can get a partial extension. I'm can't test it yet because I'm not using partial extensions and ( in contrast to before ) I decided not to add something completely different 1.5 week before a tournement. Tony > > >> >>>Uli >>> >>>> >>>>-- >>>>GCP
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.