Author: Tord Romstad
Date: 10:15:12 01/17/04
Go up one level in this thread
On January 17, 2004 at 12:40:37, Christophe Theron wrote: >On January 17, 2004 at 04:51:30, Russell Reagan wrote: > >>On January 16, 2004 at 19:06:21, Harald Faber wrote: >> >>> >>>[D]2r1k1b1/pp4Q1/3p4/q5p1/4P3/1PN2P2/PP6/1K5R b - - 0 1 >>> >>>Analysis by Shredder 8: >>> >>>1. +- (#153): 33...Lf7 34.Th8+ Ke7 35.Txc8 d5 36.Sxd5+ Kd6 37.Dxf7 De1+ 38.Tc1 >>>Dc3 39.bxc3 b5 40.Dxa7 Kc6 41.Dc7# >>>2. +- (#151): 33...Tc7 34.Dxg8+ Kd7 35.Sd5 Tc2 36.Th7+ Kc6 37.Dc8+ Dc7 38.Txc7+ >>>Kb5 39.Kxc2 a5 40.Dxb7# >>>3. +- (#7): 33...De5 34.Dxg8+ Kd7 35.Th7+ De7 36.Txe7+ Kxe7 37.Dxc8 >> >>Oops...this is easily explainable. It took a minute for me to remember this. It >>is probably a fail high in a root aspiration search. You know how sometimes you >>will see a program search like this? >> >>1. +0.25 e4 e5 Nf3 Nf6 >>2. +0.27 e4 e5 Nf3 Nf6 Bb5 >>3. ----- d4! <--- This is the line of interest >>3. +0.45 d4 d5 c4 e6 >> >>The line where there is no score and no PV usually means that a root aspiration >>search was done and the search failed high. Most programs don't give a PV, or >>don't give a score, or don't give either. Shredder probably just gives both. >> >>The reason is that the fail high in the aspiration search just means that this >>new move should be better, but you don't have an exact score yet, and you have >>to re-search with a full alpha-beta window to get the exact score. It is >>possible that #153 and #151 are the inaccurate scores returned by a fail high >>aspiration search at the root, and they are resolved to exact scores later. > > > >I think you have found the correct explanation. > >Chess Tiger can do that at times. It finds a forced mate with an inaccurate >score because the real score is outside the window. But that inaccurate score is >interpreted by the machine to human best line converter. A value close to +30000 >or -30000 is converted to a mate score (distance to mate = +30000 - score for >example). This gives an incorrect distance to mate because the score is not >correct, it's just a bound (the program knows that the real score is above >that). > >When the program resolves the fail-high, the correct score is found. In turn it >converts to a correct distance to mate. > >I think this problem is even worse with programs using MTD(f). That's entirely correct, at least for my engine. It is very normal that it reports mate in too many moves. The correct way to interpret "mate in n moves" is "mate in at most n moves". I don't consider it to be a serious problem, though. While we're talking aboug Chess Tiger and resolving fail highs, there is one aspect of Tiger's behaviour on my Palm which has puzzled me for a long time. It often happens that it searches the first move in a new iteration, reports a fail high on one of the subsequent moves, but after resolving the fail high decides to keep the first move. What is the explanation for this? Tord
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.