Author: Gordon Rattray
Date: 13:52:19 10/23/01
Go up one level in this thread
On October 23, 2001 at 16:34:35, Robert Hyatt wrote: >On October 23, 2001 at 13:05:49, Gordon Rattray wrote: > >>On October 23, 2001 at 11:37:07, Robert Hyatt wrote: >> >>>On October 23, 2001 at 10:19:53, Gordon Rattray wrote: >>> >>>>On October 23, 2001 at 09:58:52, Robert Hyatt wrote: >>>> >>>>>On October 23, 2001 at 08:56:18, Gordon Rattray wrote: >>>>> >>>>>>Suppose that during the analysis of a position, a chess engine shows evaluations >>>>>>similar to this: >>>>>> >>>>>>depth 7: Qe1 eval: 0.58 >>>>>>depth 8: Qe1 eval: 0.42 >>>>>>depth 9: Qe1 eval: 0.39 >>>>>>depth 10: Qe1 eval: 0.33 >>>>>> >>>>>>i.e. as the depth increases, the move choice stays the same, and the evaluation >>>>>>is slowly but consistently decreasing. >>>>>> >>>>>>I release that in general the next evaluation could be anything, but I am right >>>>>>in thinking that the probability of it decreasing further is higher than normal? >>>>>> Or to take it to extremes, I think the above is more likely to decrease further >>>>>>than the following: >>>>>> >>>>>>depth 7: Qe1 eval: 0.33 >>>>>>depth 8: Qe1 eval: 0.39 >>>>>>depth 9: Qe1 eval: 0.42 >>>>>>depth 10: Qe1 eval: 0.58 >>>>>> >>>>>>Did each line involve a "fail low"? Or am I getting my terms mixed up? >>>>>> >>>>>>Do chess programs take account of the above and, e.g., are more likely to search >>>>>>further before moving? >>>>>> >>>>>>Gordon >>>>> >>>>> >>>>>That generally suggests that there is a horizon-effect happening, and that >>>>>eventually the best move will be bad enough that a new best move will be >>>>>found. >>>> >>>>Thanks, that confirms my thinking. One position in which I seen the effect >>>>involed White grabbing a pawn at the expense of uncoordinated pieces. So, the >>>>term "horizon effect" seems applicable. >>>> >>>>> >>>>>But searchng until you find it is not possible in a timed game, for obvious >>>>>reasons... >>>> >>>>Sure, but can't the timing code take this into some consideration? For example, >>>>if the next iteration is estimated to take too much time, but only just over the >>>>allowable period, would it be worthwhile paying that bit extra? Alternatively, >>>>can increasing evaluations be used as a factor for moving quicker? I'm not >>>>suggesting that there wouldn't be other factors in either case. >>>> >>>> >>>>Gordon >>> >>>That is the trivial case. But suppose the _last_ iteration is only .1 worse >>>than the previous iteration. And suppose the next iteration will take 20 >>>minutes (estimated) to search? And if that one is again .1 worse, the next >>>will take probably an hour... >> >> >>Agreed. >> >>It's the probability of avoiding a possible horizon effect, versus an >>exponentially increasing time penality. It may be that the root position is >>just bad and looks worse with each additional ply. >> >>I was just curious as to whether this concept was worth considering, even for >>"trivial" cases... but how likely are they to occur... maybe not worth the >>hassle... >> >>Gordon > > >Something is doable here. Deep Blue had some sort of algorithm that would >extend the time when the tree was "unstable". And they didn't just define >unstable as "fail lows or excessive move-changing at the root." They had >something that Hsu briefly explained to me once several years ago, but there >was a lot going on and I didn't pay close enough attention to remember enough >to try it. They could recognize (at least for their program) when something >unusual was going on and use more time to try to "search through" the problem >and see what was _really_ going to happen... > >One day, I will think about it and ask them for an explanation again. This >time I'll take notes. :) :-) Interesting, thanks for your input. Gordon
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.