Author: Frank E. Oldham
Date: 14:57:27 05/03/04
Go up one level in this thread
On May 02, 2004 at 10:47:13, Robert Hyatt wrote:
>On May 01, 2004 at 19:33:02, J. Wesley Cleveland wrote:
>
>>On May 01, 2004 at 17:33:34, Vincent Lejeune wrote:
>>
>>>On May 01, 2004 at 17:05:27, J. Wesley Cleveland wrote:
>>>
>>>>On April 30, 2004 at 23:43:48, Robert Hyatt wrote:
>>>>
>>>>>On April 30, 2004 at 17:21:34, José Antônio Fabiano Mendes wrote:
>>>>>
>>>>>> [D]7k/6p1/2P3Qp/p3q2P/8/6P1/5K2/8 w
>>>>>> Bogoljubow vs Stahlberg, Kemeri 1933 1. Qc2!! Qxh5 2. Qc4!
>>>>>
>>>>>This one takes crafty depth=16, not real quick. Almost 4 minutes to pick Qc2,
>>>>>several more to see how good it really is...
>>>>>
>>>>[snip]
>>>>While looking at this with crafty, I found the ugliest fail high I've seen with
>>>>crafty.
>>>>
>>>>[D]2q4k/6p1/2P4p/p7/2Q5/6P1/6K1/8 w - - 0 4
>>>>
>>>>hash table memory = 96M bytes.
>>>>pawn hash table memory = 6M bytes.
>>>>
>>>>Crafty v19.10
>>>>
>>>>White(1): sd 11
>>>>search depth set to 11.
>>>>White(1): 2q4k/6p1/2P4p/p7/2Q5/6P1/6K1/8 w - - 0 4
>>>>White(1): analyze
>>>>Analyze Mode: type "exit" to terminate.
>>>>end-game phase
>>>> clearing hash tables
>>>> time surplus 0.00 time limit 22.50 (3:30)
>>>> depth time score variation (1)
>>>> 7-> 0.12 -0.14 1. Qb5 Qd8 2. Kf3 Qf8+ 3. Kg4 Qd6 4.
>>>> Kf5 Qc7
>>>> 8 0.18 -0.15 1. Qb5 Qd8 2. Kf3 Qd6 3. Qc4 Qf6+ 4.
>>>> Kg4 g6 5. Qd5
>>>> 8 0.46 -0.03 1. Kf3 Qd8 2. Qe4 Qd1+ 3. Kf4 Qd6+
>>>> 4. Kf3 Qc7 5. Qd5 <HT>
>>>> 8-> 0.48 -0.03 1. Kf3 Qd8 2. Qe4 Qd1+ 3. Kf4 Qd6+
>>>> 4. Kf3 Qc7 5. Qd5 <HT>
>>>> 9 0.95 -0.05 1. Kf3 Qf5+ 2. Ke3 Qe5+ 3. Kf3 Qf6+
>>>> 4. Ke4 Qe7+ 5. Kf3 Qa3+ 6. Kg2 Qf8
>>>> 7. c7 Qa8+ 8. Kf2 Qc8
>>>> 9 1.14 0.01 1. Qd5 a4 2. Qd7 Qa6 3. c7 Qe2+ 4.
>>>> Kh3 Qh5+ 5. Kg2 Qe2+
>>>> 9-> 1.15 0.01 1. Qd5 a4 2. Qd7 Qa6 3. c7 Qe2+ 4.
>>>> Kh3 Qh5+ 5. Kg2 Qe2+
>>>> 10 1.28 0.01 1. Qd5 a4 2. Qd7 Qa6 3. c7 Qe2+ 4.
>>>> Kh3 Qh5+ 5. Kg2 Qe2+
>>>> 10 1.87 ++ 1. c7!!
>>>> 10 17:13 1.45 1. c7 Qb7+ 2. Kh2 Qc8 3. Qc5 Kh7 4.
>>>> g4 Kh8 5. Qxa5 <HT>
>>>> 10-> 17:13 1.45 1. c7 Qb7+ 2. Kh2 Qc8 3. Qc5 Kh7 4.
>>>> g4 Kh8 5. Qxa5 <HT>
>>>> 11 17:14 ++ 1. c7!!
>>>> 11 17:15 5.72 1. c7 Qb7+ 2. Kh2 Qc8 3. Qc5 Kh7 4.
>>>> g4 a4 5. Qf5+ Qxf5 6. gxf5 a3 7. c8=Q
>>>> a2 8. Qc3
>>>> 11-> 17:15 5.72 1. c7 Qb7+ 2. Kh2 Qc8 3. Qc5 Kh7 4.
>>>> g4 a4 5. Qf5+ Qxf5 6. gxf5 a3 7. c8=Q
>>>> a2 8. Qc3
>>>> time=17:15 cpu=88% mat=-1 n=1791242494 fh=99% nps=1.73M
>>>> ext-> chk=381633369 cap=48652 pp=55417 1rep=27688883 mate=994
>>>> predicted=0 nodes=1791242494 evals=48417627
>>>> endgame tablebase-> probes=0 hits=0
>>>> hashing-> 65%(raw) 64%(depth) 99%(sat) 99%(pawn)
>>>> hashing-> 0%(exact) 56%(lower) 1%(upper)
>>>
>>>
>>>Shredder think it's a draw after 1.c7, some pertpetual missed to crafty I think.
>>>Could you let crafty think a little longer ? I'm about sure you will see the
>>>"ugliest fail low you've seen with crafty" ;)
>>>
>>Crafty fails low to a draw score quickly on the next ply. What makes this so
>>ugly is that it takes 1.87 seconds to fail high, and 17:13 *minutes* to resolve
>>the fail high.
>
>
>You should be using 19.12 which doesn't do this as it fails high/low in
>increments, not all at once.
using crafty19.10f
removing the sd=11 restriction, with one processor I can duplicate the fail-high
then -low:
10 2.22 ++ 1. c7!!
10-> 13:07 0.40 1. c74
11 13:07 ++ 1. c7!!
11-> 13:46 0.79 1. c78+
12 13:47 ++ 1. c7!!
12-> 13:48 1.18 1. c78+
13 13:50 -- 1. c7
13 13:59 0.01 1. c7 a4 2. Qc5 Kh7 3. Qe7 a3 4. Qd8
Qb7+ 5. Kh2 Qb2+ 6. Kh3 Qb1 7. c8=Q
Qh1+ 8. Kg4 Qe4+ 9. Kh3 Qh1+
but re-testing, with two processors, the problem doesn't appear:
11 2.11 ++ 1. c7!!
11-> 2.21 0.40 1. c7 (s=5)
12 2.57 ++ 1. c7!!
12-> 2.73 0.79 1. c7 (s=2)
13 3.30 -- 1. c7
13 12.00 0.01 1. c7 a4 2. Qc5 Kh7 3. Qa7 Qg4 4. Qb7
Qe2+ 5. Kh3 Qh5+ 6. Kg2 Qe2+
so 9 mins to resolve the fail-high for mt=1 turns into
just 10 secs for the full fail-high then -low resolution with mt=2 --
interesting!
Frank
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.