Designing an analysis friendly Stockfish?

[Jeremy Bernstein]

Hi Jeremy

Sorry for my choice of words, but apparently this has been said before ( the time losses), and nothing has been done to fix this.
In Fritz Gui the engine losses quite often on time.

And normal Stockfish never? I'll look into this, but I'm at a bit of a loss as to what it could be. Unless Stockfish's time management code makes some assumptions about the amount of time required to search a move which are somehow incompatible with the tablebase access. Or maybe there's a bug. No idea. I'll put it on my list for the next rainy day.

Jeremy

[keoki010]

Minimum Thinking Time = 2
Emergency Base Time = 20
Emergency Move Time = 7

Ran a match using these settings on Gran2c and Gran2d and still some time losses in 3"+3'. I'm going to try normal 201 ja against Gran2d and see what happens.

[Damir Desevac]

Thanks Jeremy. It would really be great if you can fix this time bug, which occurs quite often, no matter what time settings I use.

[keoki010]

Ran a new tour only 51 games but no time losses. Used the above values. Will run another 51 games tonight. 3/3 and 1024 hash.

Program Elo + - Games Score Av.Op. Draws

1 Stockfish_201_PA_GTB_Gran2d_x64 : 2441 72 67 34 58.8 % 2379 64.7 %
2 Stockfish_201_64_PA_G : 2386 81 82 34 47.1 % 2407 52.9 %
3 Critter_0.90_64bit_SSE4 : 2373 81 82 34 44.1 % 2414 52.9 %
Suprisingly both versions of Stockfish beat Critter. I used SilverSuite2 for the openings, random access. Computer i7-980 @3.2

[ivanhoe]

Stockfish_201_PA_GTB_Gran2c_x64 is the fastest engine out of all the current compiles.
AMD Quad 3.4GHz
4Gb DDR3 1600Mhz Ram
XP64 OS.

...and probably the strongest

[Jeremy Bernstein]

Gran2e uses a more conservative time management calculation. Running some hyperbullet games between this, Gran2d and the JA build, I am not seeing any time losses for Gran2e, and plenty for the other two, using late endgame positions as a starting point. Give it a whirl, but no promises.

If it doesn't work better, maybe someone could give me a step-by-step for creating time losses using a ChessBase product. I am not convinced that Arena is properly adjudicating time losses -- at least the interface displays very strange stuff most of the time (negative times for instance).

Jeremy

[Jeremy Bernstein]

Engine                          Score                                                                                                                                                                                                          St                                                                                                                                                                                                       St                                                                                                                                                                                                       St    S-B
Stockfish_201_PA_GTB_Gran2e_x64 216,0/400 ········································································································································································································ =====111=0==01=1=11001=010==10====1=01==01010111==110101001011==0111==0101==010110========011===01==0101==1===1=01010===1001====01=0======111001010101==1===01==10100101101=01=110==1010100====001==1001 =001====1000011=000110=0====0101===1101=0===0110=10=0110111=0011101=110=11=1=01011100111==01=11001==10=1011=01=11=1010==10==1000==1001==00101110====10101=111011==1001100=111010101010101101======0110==  41434,
Stockfish_201_PA_GTB_Gran2d_x64 211,0/400 =====000=1==10=0=00110=101==01====0=10==10101000==001010110100==1000==1010==101001========100===10==1010==0===0=10101===0110====10=1======000110101010==0===10==01011010010=10=001==0101011====110==0110 ········································································································································································································ 1===01110111010101101011==11101101010101=11=11======01=1100=0111==010101010111101101010110=001=1==01=001==0111==010101011=101110010=0110111==0011110==110=010101==01101001=1====0101=1111001=0=11=01==11  40502,
Stockfish-201-64-ja             173,0/400 =110====0111100=111001=1====1010===0010=1===1001=01=1001000=1100010=001=00=0=10100011000==10=00110==01=0100=10=00=0101==01==0111==0110==11010001====01010=000100==0110011=000101010101010010======1001== 0===10001000101010010100==00010010101010=00=00======10=0011=1000==101010101000010010101001=110=0==10=110==1000==101010100=010001101=1001000==1100001==001=101010==10010110=0====1010=0000110=1=00=10==00 ········································································································································································································  36958,

600 games played / Tournament is finished
Level: Blitz 0:05/0,1
Hardware: Intel(R) Xeon(R) CPU  W3520 @ 2.67GHz 2659 MHz with 1,0 GB Memory
Operating system: Windows 7 Professional Professional (Build 7600) 64 bit

Note that these were all endgame positions, so the GTB builds have a built-in advantage where the position is winnable or drawable. Since many of the positions were forced WLD, it's hard to evaluate which engine is actually stronger. PGNs are attached.

[Jeremy Bernstein]

By the way, running Gran2e v JA right now in Fritz 1/1: both engines lose on time, much of the time. There's something up with the CB interface (at least on a 32-bit 1-core system) where the engines just sit and stall for a few seconds from time to time and then suddenly dump some data into the engine window. This might have something to do with what I'm seeing, so maybe someone with a mightier system wants to check that out a little closer.

Anyway, I am prepared to say that the TM is equivalent between the release and the modified versions of Stockfish. Whether that means that standard Stockfish (also) has "crap time management" is a question for the curious reader. In any case, the tablebase access code isn't making it any worse.

For the archives, Gran2e is very careful about not wasting time with tablebase lines if more than ~6% of the available search time has been used up. Gran2d was more generous (~60% was permitted before capping off the probing).

Jeremy

[keoki010]

By the way, running Gran2e v JA right now in Fritz 1/1: both engines lose on time, much of the time. There's something up with the CB interface (at least on a 32-bit 1-core system) where the engines just sit and stall for a few seconds from time to time and then suddenly dump some data into the engine window. This might have something to do with what I'm seeing, so maybe someone with a mightier system wants to check that out a little closer.

Anyway, I am prepared to say that the TM is equivalent between the release and the modified versions of Stockfish. Whether that means that standard Stockfish (also) has "crap time management" is a question for the curious reader. In any case, the tablebase access code isn't making it any worse.

For the archives, Gran2e is very careful about not wasting time with tablebase lines if more than ~6% of the available search time has been used up. Gran2d was more generous (~60% was permitted before capping off the probing).

Jeremy

Have you tried setting the TC management parameters I have in a post above. With 102 3/3 games I had no time losses in Arena. I'll try with Probe the root not checked and see if that had anything to do with it.

[keoki010]

Latest tour. ponder=off defauts for all but Gran2d. NO time losses 3/3




Program Elo + - Games Score Av.Op. Draws

1 Stockfish-201-64-ja : 2441 62 49 34 58.8 % 2379 76.5 %
2 Stockfish_201_PA_GTB_Gran2d_x64 : 2400 64 64 34 50.0 % 2400 70.6 %
3 Stockfish_201_PA_GTB_Gran2e_x64 : 2359 49 62 34 41.2 % 2421 76.5 %
Individual statistics:

1 Stockfish-201-64-ja : 2441 34 (+ 7,= 26,- 1), 58.8 %

Stockfish_201_PA_GTB_Gran2d_x64: 17 (+ 5,= 12,- 0), 64.7 %
Stockfish_201_PA_GTB_Gran2e_x64: 17 (+ 2,= 14,- 1), 52.9 %

2 Stockfish_201_PA_GTB_Gran2d_x64: 2400 34 (+ 5,= 24,- 5), 50.0 %

Stockfish_201_PA_GTB_Gran2e_x64: 17 (+ 5,= 12,- 0), 64.7 %
Stockfish-201-64-ja : 17 (+ 0,= 12,- 5), 35.3 %

3 Stockfish_201_PA_GTB_Gran2e_x64: 2359 34 (+ 1,= 26,- 7), 41.2 %

Stockfish_201_PA_GTB_Gran2d_x64: 17 (+ 0,= 12,- 5), 35.3 %
Stockfish-201-64-ja : 17 (+ 1,= 14,- 2), 47.1 %