Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: crafty fail high problem was Re: A really nasty position

Author: Robert Hyatt

Date: 12:14:46 05/30/01

Go up one level in this thread


On May 30, 2001 at 13:23:57, J. Wesley Cleveland wrote:

>On May 29, 2001 at 17:19:09, Robert Hyatt wrote:
>
>>On May 29, 2001 at 14:21:13, Peter Fendrich wrote:
>>
>>>This position appeared in an ICC game: Guitje - TerraPi
>>>
>>>[D]2r3kb/5q1p/4pPp1/1p1pP3/p2P2Q1/PrP5/1P3R1R/2B3K1 w - - 0 42
>>>
>>>the last move from Terra was 41... b5
>>>Some moves earlier when white made g6, Terra decided to move the bishop from g7
>>>to h8.
>>>Both Rb3 and Bh8 are locked in and Terra is absolutely unaware of the
>>>hopeless situation. Yes, Terra lost...
>>>
>>>I see two different but connected problems for computer programs here:
>>>  - the closed position itself
>>>  - the two blocked pieces.
>>>
>>>How do you solve/avoid positions like this? Terra has about equal eval for this
>>>position.
>>>It is of course possible to punish pawn chains like this in order to avoid
>>>closed positions but is that really a solution and how to deal with blocked
>>>positions?
>>>
>>>//Peter
>>
>>
>>First, this looks dead lost.  I get big scores (>3) after a few seconds and 13
>>plies.
>>
>>The position you gave is too late to repair.  You have to recognize blocked
>>pawns _before_ they are blocked, which means the 'mistakes' were made earlier
>>in the game.  I evaluate blocked pawns in Crafty, but I also evaluate pawn
>>levers and make sure I keep a pawn break opportunity alive to avoid getting
>>into totally closed positions...
>>
>>Here you should probably find Rxh7 is crushing of course...
>
>I ran this position on crafty 18.3 and got this log. Between plies 13 and 14
>there is an effective branching factor of almost 600. I know that fail highs can
>cause big problems, but this seems extreme.
>
>               13    27.10   3.77   1. Rxh7 Kxh7 2. Qh4+ Kg8 3. Rh2 Bxf6
>                                    4. exf6 Kf8 5. Qh8+ Qg8 6. Rh7 g5 7.
>                                    f7 Qxh8 8. Rxh8+ Kxf7 9. Rxc8 g4 10.
>                                    Kf2 Kf6
>               13->  34.05   3.77   1. Rxh7 Kxh7 2. Qh4+ Kg8 3. Rh2 Bxf6
>                                    4. exf6 Kf8 5. Qh8+ Qg8 6. Rh7 g5 7.
>                                    f7 Qxh8 8. Rxh8+ Kxf7 9. Rxc8 g4 10.
>                                    Kf2 Kf6
>               14    43.34     ++   1. Rxh7!!
>               14   335:00   4.30   1. Rxh7 Kxh7 2. Qh4+ Kg8 3. Rh2 Bxf6
>                                    4. exf6 Kf8 5. Qh8+ Qg8 6. f7 Qxh8
>                                    7. Rxh8+ Kxf7 8. Rxc8 Kf6 9. Kg2 Kf5
>                                    10. Rf8+ Ke4 11. Rf4+ Kd3 12. Rf3+
>                                    Ke4 13. Kg3
>               14-> 335:22   4.30   1. Rxh7 Kxh7 2. Qh4+ Kg8 3. Rh2 Bxf6
>                                    4. exf6 Kf8 5. Qh8+ Qg8 6. f7 Qxh8
>                                    7. Rxh8+ Kxf7 8. Rxc8 Kf6 9. Kg2 Kf5
>                                    10. Rf8+ Ke4 11. Rf4+ Kd3 12. Rf3+
>                                    Ke4 13. Kg3


This is probably the result of lots of things.  1.  hash too small.  at 1M
nodes per second, you just searched 335*60*1M nodes.  it probably was like
searching with no hash at all;  2.  a fail high means something unexpected
happened.  Move ordering is therefore shot to hell since what you thought
was right is now wrong and vice-versa.  3.  When I fail high, I just relax
beta to +infinity, rather than the tiered approach I used in Cray Blitz.  If
there are lots of mates here, that are not forced, the program still has to
search them all out as with +infinity, you get zero beta cutoffs at positions
where white is to move, until you establish a better beta value after a lot
of searching.




This page took 0.01 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.