Author: Will Singleton
Date: 20:40:36 03/08/04
Go up one level in this thread
On March 08, 2004 at 14:20:24, Robert Hyatt wrote: >On March 08, 2004 at 13:20:26, J. Wesley Cleveland wrote: > >>On March 07, 2004 at 13:47:03, Robert Hyatt wrote: >> >>>On March 06, 2004 at 22:20:32, Will Singleton wrote: >>> >>>>On March 06, 2004 at 20:45:17, scott farrell wrote: >>>> >>>>>On March 06, 2004 at 20:02:13, GeoffW wrote: >>>>> >>>>>>Hi Dan >>>>>> >>>>>>Thanks for the extensive analysis, that was a lot of output so I summarised the >>>>>>times to avoid the obvious Bishop takes rook >>>>>> >>>>>> >>>>>>Amy-net-087............00:16 >>>>>>Aristarch 4.41.........00:02 >>>>>>DeepSjeng..............00:04 >>>>>>Delfi-440..............00:13 >>>>>>Dragon_45..............01:23 >>>>>>ELChinito 3.25.........00:06 >>>>>>Glc300.................00:05 >>>>>>Gothmog................01:31 >>>>>>Kke-253................00:08 >>>>>>Ktulu..................00:23 >>>>>>List512................00:04 >>>>>>Patzer 3.61.......... >00:35 not resolved >>>>>>Quark-232-net..........00:53 >>>>>>Ruffian_202............00:03 >>>>>>Ruffian_210........... 00:04 >>>>>>Smarthink-017a.........00:10 >>>>>>Yace...................01:06 >>>>>>Shredder 7.04..........00:03 >>>>>> >>>>>> >>>>>>This is seemingly a good test position to separate the elite programs from the >>>>>>merely good ones, the real top programs can see and resolve this threat in less >>>>>>than 5 seconds. >>>>>> >>>>>>I have only got to improve my program by a factor of 300 to get there, Arrrrrgh >>>>>>! >>>>>> >>>>>>It looks as though my program is the only one that really suffers from a nodes >>>>>>explosion though. >>>>> >>>>>Instead of trying to solve the fail low at depth 12, and cause a node explosion, >>>>>on the fail low - go back to either depth 11 to reasearch, or depth 1. This way >>>>>the hash will know that the captures will fail, and start choosing seemingly >>>>>poorer moves at lower depths, but the fail low will be solved in many less >>>>>nodes. After seeing the faillow at depth 12, it might solve it using the hash at >>>>>depth 9. >>>>> >>>>>Scott >>>>> >>>> >>>>I do something like that, and I think it can help for some score drops. (I >>>>don't have fail-lows, I have score drops at the root (pvs, window wide open or >>>>closed).) Note that the ply changes from 12 to 9 after the bad score, allowing >>>>a new move selection in a shorter time. >>>> >>>>8 191 20 62537 Bxf6 exf6 g6 Bf2 Bc6 Qg4 Qd8 >>>>9 187 35 115687 Bxf6 gxf6 g6 hxg6 >>>>10 173 49 169855 Bxf6 gxf6 g6 hxg6 fxg6 Bf2 Qd8 >>>>11 131 95 334767 Bxf6 gxf6 g6 Qg5 Nd7 hxg6 fxg6 Nxg6 hxg6 Bxg6 Qd8 >>>>12 -293 646 2281714 Bxf6 gxf6 Ng6 hxg6 fxg6 Nxg6 Bc6 Ne7 Rxe7 Qh4 h5 >>>>9 -293 646 2281715 Bxf6 >>>>9 131 765 2698812 b4 >>>>10 105 890 3143214 b4 Rf3 bxa3 bxa3 Bxa3 h6 g6 Rf1 Bc6 Bf2 >>>>11 105 1162 4192271 b4 axb4 axb4 Rf4 b3 >>>>12 99 1968 7109127 b4 axb4 Bxf6 exf6 g6 hxg6 hxg6 bxa5 Rxa5 >>>>13 106 3434 12510952 b4 axb4 axb4 Rf4 Ra2 Qg2 b3 Bb1 Ra1 Bf2 Nd7 h6 g6 >>>> >>>>That's amateur 2.80 on a p4 2.8ghz. >>>> >>>>Will >>> >>> >>>here is a technical issue question: >>> >>>you fail low on move x. you back up 4 plies. what do you do if move x is >>>searched first and it fails low instantly??? back up 4 more? repeat? etc? >>> >>>I found problems trying this in Crafty a good while back. As in a loop that >>>won't end until time runs out... >> >>I think you misread what he is doing. If he gets the equivalent of a fail low at >>the root at ply n, he widens the search window and restarts the search at ply >>n-3(a form of iterative deepening). I tried something similar with Crafty and it >>helped quite a bit on fail lows. The problem is that after a fail low, the hash >>table is of little use, esp. with failsoft, and move ordering goes to pot. > > >Perhaps you missed my point here... > >If you reduce the depth in Crafty, _how_ do you guarantee that when you back up >N plies and start the search, you don't get the fail-low move sorted first to >search? If you do, you die. It will _always_ fail low and you enter a loop >that won't terminate until time runs out. That is the problem here, at least in >Crafty. you get an infinite loop that will eventually have to play the very move >that failed low... Not what you intended... > >This is a problem because the oroginal hash info on the fail low will cause any >re-search to a shallower depth to instantly fail low... and there starts the >loop. > >BTW... I have seriously hurt my right hand, and as a result, I can't use it for >at least four weeks as it recovers from surgery. As a result, my explanations >will likely be a bit short and/or cryptic for obvious reasons... But I'll >still try to respond as necessary to questions, but typing/programming will be >very slow for a while. :( Sorry to hear about your hand. As a matter of fact, I went paintballing with my son on Saturday, and I'm still sore all over. :) As I mentioned, my approach to score-drops is a kludge that probably should be handled in a more enlightened way. But to answer your question, I don't get repeated score drops because on the reduced ply re-search (plys 9-12 in the above example), the ab window is wide open for the first move, and the score for the offending root move is stored and returned as is without a search, until the ply depth exceeds the original score-drop ply. It's possible that a different root move will again cause a score drop on the re-search, but in practice it's not likely, and I also limit these re-searches to 3 times per search, as well as reduce the reduction by 1 each time. Will
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.