What makes 1-cpu program un-deterministic?

[Jouni]

Obviously all SMP software un-deterministic, but why are some 1-cpu programs also un-deterministic. Specially Critter is totally un-predictable even in single version! Has Richard added some randomness in search?

Jouni

[rvida]

Obviously all SMP software un-deterministic, but why are some 1-cpu programs also un-deterministic. Specially Critter is totally un-predictable even in single version! Has Richard added some randomness in search?

Jouni

It is supposed to be deterministic. If it is not you probably found a bug. With a freshly started engine, outputs should be identical even across different platforms.


Richard

[WinPooh]

Probably, cosmic rays?

[hgm]

Any program that uses the clock to make search decisions will be non-deterministic.

[WinPooh]

Not cleaning up hash space or killers/history datastuctures on startup can lead to nondeterministic behaviour too.

[rvida]

Any program that uses the clock to make search decisions will be non-deterministic.

I assumed the question was about analysis, not gameplay.

[hgm]

Well, waiting for user moves is even more unpredictable than waiting for the clock to tick. Or do you clear the hash table before each search?

[ernest]

. Specially Critter is totally un-predictable even in single version!

I think something must be wrong with your tests...
(do you clear hash? )
Can you give an example?

In Fritz GUI,using 1 thread (512MB hash), I always get the same analysis from the start_position:

Code: Select all

rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
Analysis by Critter 1.2 32-bit gtb:
...............
1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Bc4 Bc5 5.0-0 0-0 6.d3 d6 7.Be3 Nd4 8.h3 Be6 9.Nxd4 exd4 
  =  (0.15 --)   Depth: 18/37   00:00:10  11730kN
1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Bc4 Bc5 5.0-0 0-0 6.d3 d6 7.a3 h6 8.b4 Bd4 9.Nxd4 Nxd4 10.h3 Be6 
  =  (0.13)   Depth: 18/37   00:00:11  13682kN
1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Bc4 Bc5 5.0-0 0-0 6.d3 d6 7.Na4 Be6 8.Nxc5 dxc5 9.Bxe6 fxe6 10.Be3 Qd6 11.c3 Rad8 12.h3 
  =  (0.14)   Depth: 19/40   00:00:21  24104kN
1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Bc4 Bc5 5.0-0 0-0 6.d3 d6 7.Na4 Be6 8.Nxc5 dxc5 9.Bxe6 fxe6 10.Be3 Qd6 11.c3 Rad8 12.h3 
  =  (0.14)   Depth: 20/40   00:00:33  38635kN
1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Be2 Bd6 5.0-0 0-0 6.d3 b6 7.a3 h6 8.Be3 Ng4 9.Qd2 Nxe3 10.fxe3 Bb7 11.Nd5 Ne7 
  =  (0.11)   Depth: 21/43   00:01:10  81550kN
1.d4 Nf6 2.Nc3 d5 3.Bf4 e6 4.e3 Be7 5.h3 0-0 6.Nf3 Ne4 7.Bd3 Nxc3 8.bxc3 c5 9.0-0 c4 10.Be2 Nc6 11.Ne5 Bd6 12.Rb1 Nxe5 13.Bxe5 Bxe5 14.dxe5 
  =  (0.12)   Depth: 21/43   00:02:14  154mN

[Jouni]

I used/wasted 3 hours to re-test everything. And yes 1-cpu Critter is clearly deterministic (sorry Richard), but Chessbase GUI (12) isn't When running testsuite in Arena You get same result in second run. In Chessbase GUI there is difference (hash clearing / engine loading thing?) and results vary a lot with Critter! But note: Rybka and Stockfish are totally deterministic even in Chessbase GUI = there is difference..

Jouni

[ernest]

Don't know about Chessbase GUI (12). Do you mean Fritz 12 GUI?

I use Fritz 11 GUI.