Author: Uri Blass
Date: 23:16:14 09/27/01
Go up one level in this thread
On September 28, 2001 at 02:09:11, Uri Blass wrote: >On September 27, 2001 at 23:42:31, Robert Hyatt wrote: > >>On September 27, 2001 at 17:48:32, Peter Fendrich wrote: >> >>>On September 27, 2001 at 15:12:43, Roy Eassa wrote: >>> >>>>On September 27, 2001 at 12:13:10, Peter Fendrich wrote: >>>> >>>>>On September 26, 2001 at 21:45:46, Robert Hyatt wrote: >>>>> >>>>>>On September 26, 2001 at 20:32:58, Dann Corbit wrote: >>>>>> >>>>>-- snip -- >>>>> >>>>>>If you correctly predict your opponent's move at least 50% of the time, or >>>>>>more, then the way we currently ponder can _not_ be improved on. >>>>> >>>>>I don't agree if that's what you really mean. "can _not_ be..." is hard to prove >>>>>in this case. In theory at least you can do better. The _average_ hit rate is >>>>>>50% >>>>>If you know that this hit rate vary with different circumstances you will find >>>>>out different hit rates. If we could separate out cases with very low hit rate >>>>>it might be succesful with another scheme for just these cases. I've never >>>>>tested this but it would be interesting to see the hit rate for "consistent" >>>>>FH's (survives several iterations) compared to the rest. The hit rate for >>>>>pondermoves giving about the same evaluation as before is probably higher (much >>>>>higher?). >>>>>I can think of other types of cases as well. >>>>>Has anyone computed the figures for different cases like this? >>>>> >>>>>I would like leave this "can _not_ be..." open until at least some test like >>>>>this is done. >>>>> >>>> >>>> >>>>The factor that causes the engine to be unsure of the move it selected to >>>>ponder, is the SAME factor that makes pondering multiple moves less useful. >>>> >>>>If there are several moves that are all about equal, then there are, by >>>>definition, also several moves among which you must divide your time pondering. >>>>Thus even if you were only 20% sure of your opponent's move, it still does not >>>>make sense to split your pondering time because each likely move would then get >>>>no more than that same 20%. >>> >>>Yes, I buy all that. My intention was to oppose to the "it's impossible" >>>statement. You are talking about some general case. There is no reason why each >>>move has to be 20% because the first one is. That's why I'm talking about >>>isolating cases where the other move might be better. Another question is what >>>happens if the ponder move has only 10% or 5% probability. >>>I have no proofs that these cases are possible to identify but I'm still open >>>for it, until I know better... >>>//Peter >> >> >> >>The question is, what would cause that 10%. IE this is all speculation since >>we won't know whether the opponent will match or not, until he makes a move... >> >>But based on collected statistics, Crafty _always_ predicts at well over 50% >>accuracy. And as long as that is possible, I don't see any way possible to >>better utilize pondering time. Because it will _always_ be right over 50% of >>the time and save that time. >> >>Here is a test scenario: >> >>1. Assume my opponent _never_ predicts my moves correctly. IE crafty is the >>only one that ever predicts a move. In this case, crafty is the _only_ player >>that will save any time pondering. >> >>2. Assume Crafty predicts correctly 60% of the time, and the game being played >>is such that it has one minute per move, fixed, to make it simple. Then it >>will average saving 36 seconds per move over the game, based on that 60% >>prediction rate (.60 * 60 seconds). >> >>Now, given those two constraints, give me an algorithm that will save more >>than 36 seconds per move, on average... You can assume anything you want, >>just so you don't violate the 60% prediction rate already given. > >No problem > >Suppose that you have an algorithm to tell you in 10% of the cases that the >probability to ponder correctly is only 1%(I do not know about an algorithm to >do it but it does not mean that there is no algorthm to do it) > >It does not violate the 60% prediction rate because you may have probability of >almost 70% to predict the corect move in the rest of the cases. >In this case it may be better to ponder the root move in 10% of the cases. > > >I think that you can evaluate the probability that you ponder correctly based >on the move that you ponder better. > >For example I guess that the prediction rate when you predict waste tempo move >is smaller than the normal probability but I do not think it is something near >1%. > >If the last move was Ra1-b1 and crafty ponders on Rb1-a1 or Rb1-c1 then I guess >that the prediction rate is lower than the normal prediction rate. > >I still believe that it is high enough in this case in order not to ponder on >the root move but it is possible that with more complicated conditions you can >find cases when the probability to ponder correctly is small enough so it is a >bad idea to ponder on the opponent's move. > >Uri other cases when the probability to ponder correctly is smaller than the normal probability to ponder is when crafty changed it's opinion about the move to ponder in the last ply. if Crafty said 1.e4 e5 at ply 10 and 1.e4 c5 at ply 11 then you can be less sure that you ponder correctly relative to the case when crafty said 1.e4 e5 at ply 10 and at ply 11. I do not think that the change in the probability is high enough to justify not pondering the expected move in this case but it is possible that if you find more conditions that reduce the probability of pondering correctly you can find cases when it is a good idea not to ponder on the expected move. Uri
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.