Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: New computer chess book now available from Gambit

Author: Russell Reagan

Date: 10:26:34 05/27/04

Go up one level in this thread


On May 27, 2004 at 11:15:27, Robin Smith wrote:

>Thanks Will. I have no idea if Shredder is mtd(f) or not. As I'm sure you know,
>commercial authors are pretty tight lipped about the inner workings of their
>programs. Shredder does have the most bizzare pv's I have ever seen, and I have
>heard no explanations other than hash table issues to account for this.

Another idea comes to mind also (actually a variation on the given theme).
Sometimes an engine fails high and has discovered a new best PV, but it hasn't
resolved the fail high and found out the complete new best PV or the new best
score. It only knows the new best move. Sometimes we see Fritz (and many other
engines) give analysis like this.

[D]2qr2k1/1p3pp1/p3bn1p/3p4/Q1r5/2NR2P1/PP2PP1P/3R2K1 w - - 0 1

Analysis by Fritz 8:

1.Qa5 Rd6 2.Rd4 Rdc6 3.e3 Qf8 4.Ne2
  -+  (-3.34)   Depth: 7/21   00:00:00  88kN
1.Qa5 Rd6 2.Rd4 b5 3.e3 b4 4.Rxc4 Qxc4 5.Rd4
  -+  (-3.37)   Depth: 8/22   00:00:00  203kN
1.Qb3!
  -+  (-3.34)   Depth: 8/23   00:00:00  258kN
1.Qb3--
  -+  (-3.62)   Depth: 9/23   00:00:00  424kN
1.Qb3 b5 2.Rd4
  -+  (-3.62)   Depth: 9/23   00:00:00  646kN
1.Qa5!
  -+  (-3.59)   Depth: 9/23   00:00:00  754kN
1.Qa5!
  -+  (-3.44)   Depth: 9/25   00:00:00  857kN
1.Qa5 Rd7 2.Qb6 Qc6 3.Qa7 Rd8 4.Qe3 Re8 5.Rd4 b6
  -+  (-3.47)   Depth: 10/27   00:00:02  1918kN
1.Qa5 Rd7 2.Rd4 Ne4 3.Nxe4 dxe4 4.Rxd7 Bxd7 5.Qd2 Be6 6.Qe3 Qc7
  -+  (-3.53)   Depth: 11/28   00:00:04  4728kN

Notice these variations:

1.Qa5!
  -+  (-3.59)   Depth: 9/23   00:00:00  754kN
1.Qa5!
  -+  (-3.44)   Depth: 9/25   00:00:00  857kN
1.Qa5 Rd7 2.Qb6 Qc6 3.Qa7 Rd8 4.Qe3 Re8 5.Rd4 b6
  -+  (-3.47)   Depth: 10/27   00:00:02  1918kN
1.Qa5 Rd7 2.Rd4 Ne4 3.Nxe4 dxe4 4.Rxd7 Bxd7 5.Qd2 Be6 6.Qe3 Qc7
  -+  (-3.53)   Depth: 11/28   00:00:04  4728kN

What is happening here is that Fritz failed high with Qa5, i.e. it knew Qa5 was
the new best move, but it didn't know the remaining moves in the PV, or even the
exact score. A few seconds later, it resolves the fail high and gives a complete
PV and new scores.

Shredder may be doing the same thing, but it may be spitting out more PV moves,
even though they are not known yet. It extracts whatever is in the hash table
and spits it out. The thing that makes me think this might be the culprit is
that Shredder (like a few other engines I've seen, such as Ruffian) will give
analysis of 'mate in 180' or 'mate in 500' or something ridiculous that it
obviously didn't calculate out. This is a fail high situation where Shredder has
found a mate, but it doesn't know the exact distance yet (just like it didn't
know the exact score in the fail high above). It has to resolve the actual score
and PV in both cases. The only difference in these two examples is that with the
first, we humans have no idea if the "real" score is -3.59, -3.44, -3.47, or
-3.53 (nor do we really care in this case). But we do know that Shredder didn't
calculate mate in 180, and that tells me it's failing high and hasn't resolved
the true score yet.

Here is a thread that discusses the topic.

http://chessprogramming.org/cccsearch/ccc.php?find_thread=342912



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.