Hamsters 0.7.1 released

[Alessandro Scotti]

Interesting... I did not run the node count contest between 0.7.1 and 0.6, will try it tomorrow.
In the meanwhile I think I have found a few other bugs and maybe will get a little improvement in 0.7.2 SMP. I'm actually spending as much time as possible on SMP in this period because I feel bad when it doesn't work.
But basically there is not much difference between 0.6 and the 0.7 family, the two most important changes are:
- LMR does not use history counters in 0.7
- no more single reply extensions in 0.7
There are several bug fixes though, which collectively should contribute a few elo.
The node count should be true in 0.7, each thread counts separately then a function sums them all together.
Re: the pawn evaluation, I did pump it up a bit, but it made no difference in my tests so I just leave it in "on faith" and plan to repeat the test with more games... the score difference is not much anyway.

[Dann Corbit]

I am getting high NPS numbers now that seem roughly proportional to thread count.

I do not know what might be different, but with 2 threads I am getting 1.5 M NPS and with 4 threads I am getting 2.7 M NPS.

[Dann Corbit]

Hamsters contest with different thread count:

Code: Select all

c:\chess\winboard\Arena\Tournaments>type rating.dat


    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Hamsters 0.7.1 64 bit (4 threads) : 2785  103  98    29    69.0 %   2646   41.4 %
  2 Hamsters 0.7.1 64 bit (3 threads) : 2704  217 212     8    56.2 %   2661   37.5 %
  3 Hamsters 0.7.1 64 bit (2 threads) : 2649   95  99    30    35.0 %   2757   43.3 %
  4 Hamsters 0.7.1 64 bit (1 threads) : 2592  177 195     9    33.3 %   2712   44.4 %

[Dr.Wael Deeb]

Hamsters contest with different thread count:

Code: Select all

c:\chess\winboard\Arena\Tournaments>type rating.dat


    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Hamsters 0.7.1 64 bit (4 threads) : 2785  103  98    29    69.0 %   2646   41.4 %
  2 Hamsters 0.7.1 64 bit (3 threads) : 2704  217 212     8    56.2 %   2661   37.5 %
  3 Hamsters 0.7.1 64 bit (2 threads) : 2649   95  99    30    35.0 %   2757   43.3 %
  4 Hamsters 0.7.1 64 bit (1 threads) : 2592  177 195     9    33.3 %   2712   44.4 %

Thanks Dan,I assume that the problem maybe affects the 32 bit MP version of the program
Anyway,I'll wait for a newer release and then I'll test....

[Graham Banks]

Hamsters contest with different thread count:

Code: Select all

c:\chess\winboard\Arena\Tournaments>type rating.dat


    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Hamsters 0.7.1 64 bit (4 threads) : 2785  103  98    29    69.0 %   2646   41.4 %
  2 Hamsters 0.7.1 64 bit (3 threads) : 2704  217 212     8    56.2 %   2661   37.5 %
  3 Hamsters 0.7.1 64 bit (2 threads) : 2649   95  99    30    35.0 %   2757   43.3 %
  4 Hamsters 0.7.1 64 bit (1 threads) : 2592  177 195     9    33.3 %   2712   44.4 %

Thanks Dan,I assume that the problem maybe affects the 32 bit MP version of the program
Anyway,I'll wait for a newer release and then I'll test....

The 32-bit single-cpu Hamsters 0.7.1 is performing poorly in my CCRL testing also.

[Dr.Wael Deeb]

Hamsters contest with different thread count:

Code: Select all

c:\chess\winboard\Arena\Tournaments>type rating.dat


    Program                          Elo    +   -   Games   Score   Av.Op.  Draws

  1 Hamsters 0.7.1 64 bit (4 threads) : 2785  103  98    29    69.0 %   2646   41.4 %
  2 Hamsters 0.7.1 64 bit (3 threads) : 2704  217 212     8    56.2 %   2661   37.5 %
  3 Hamsters 0.7.1 64 bit (2 threads) : 2649   95  99    30    35.0 %   2757   43.3 %
  4 Hamsters 0.7.1 64 bit (1 threads) : 2592  177 195     9    33.3 %   2712   44.4 %

Thanks Dan,I assume that the problem maybe affects the 32 bit MP version of the program
Anyway,I'll wait for a newer release and then I'll test....

The 32-bit single-cpu Hamsters 0.7.1 is performing poorly in my CCRL testing also.

Aha,Bingo
Thanks Graham for the confirmation....
So,it looks like we've located the source of the problem and that is Hamsters 0.7.1 32 bit version is bugged....Maybe it's simply a compiler problem

[Alessandro Scotti]

This is very puzzling and I don't think it's a compile problem.
The version I have uploaded has played 1800 test games and it has definitely played stronger than 0.6. The difference is not big, but should be detectable with 3-400 games unless 0.7.1 gets really unlucky.
I have also the other big question why Hamsters underperforms in CCRL (with respect to my tests) and I'm guessing that's related to time management, which seems to be the most significant difference.
I'll be releasing one or maybe two new versions before the first week of December, trying to address those problems.
It would be really useful to me if I can have a look at your games though, otherwise it will take me ages to get enough data.

[Dann Corbit]

This is very puzzling and I don't think it's a compile problem.
The version I have uploaded has played 1800 test games and it has definitely played stronger than 0.6. The difference is not big, but should be detectable with 3-400 games unless 0.7.1 gets really unlucky.
I have also the other big question why Hamsters underperforms in CCRL (with respect to my tests) and I'm guessing that's related to time management, which seems to be the most significant difference.
I'll be releasing one or maybe two new versions before the first week of December, trying to address those problems.
It would be really useful to me if I can have a look at your games though, otherwise it will take me ages to get enough data.

Feature request:
Make thread count a UCI parameter.
The darn Arena interface has the thread count vanishing from the command line supplied for no apparent reason on occasion. (So the request is a work-around for an Arena bug).

[Alessandro Scotti]

Feature request:
Make thread count a UCI parameter.
The darn Arena interface has the thread count vanishing from the command line supplied for no apparent reason on occasion. (So the request is a work-around for an Arena bug).

It's already done! Will be in version 0.7.2 together with support for Polyglot books.

[Alessandro Scotti]

Code: Select all

    Engine               Score                                                                                                       Ha
01: Hamsters063devJ      418,5/900 .................................................................................................... 
02: AlaricWB704          62,5/100  =1=1=0011=0111=01111=1=10=110=1111=1=11=01011111=11011=0011==1==0==0=11100=010=100=111010==001=10111 
03: SlowChess Blitz WV21 60,5/100  ==11=0=10=01=011=101111=1100111111=1010==1=11101=001010==11=1101===10=001001=010110110=1=01=0111=101 
04: Glaurung 1.2.1       59,5/100  ==1=1==1111=0=111011=1==11011===0==101=1110=100=11101==00=010===100101110=1=10000010=111011=111=0011 
05: Spike10a             56,0/100  01110=1=11=111=1010111==01=000001===01=11==01=1==1101=01=01001=010011==100=1=1010010=111=10===001011 
06: Pharaon 3.5.1        53,0/100  =11=000=1110===101001=111=00001=11=10=1=010101====100=010101110001=1=11=0==11=11010==00110=0=10100=1 
07: Pseudo07c            50,5/100  10=0010101011=110==101=0=01=111====1==111111=1001001==0000===100110100===111000101110001000=0=101=00 
08: SmarThink            48,0/100  =0000==0011=0110=00101=0001101011===1001011111==011000=1==000=====0==11110=10==1=1000===1==01==00011 
09: Booot413             46,0/100  10=1001011====00010=11==0==10=01=1=001=10100=0110000===1000=0=1=110011010=1=00=10=0=====0110=1100=10 
10: Colossus2007c        45,5/100  =1=0100010=0010=001=00=01===100000=011=011=110==0=111=100=11=00==00=00=11=100101010111010=11=001=00= 

    Engine               Score                                                                                                       Ha
01: Hamsters081devF      460,5/900 .................................................................................................... 
02: Glaurung 1.2.1       61,0/100  1==11100110===11=10=01011010101111001==1=101=11=00=11=11=10011=1111=01011010110==110==1=00==011=10=1 
03: SlowChess Blitz WV21 59,0/100  110100==1==10011==10111011====11=11111000=0101=1=1===11001100=001==1=1010=01011=1100=1==11=01111=010 
04: AlaricWB704          53,5/100  1110111=1110=101101===00000=110===100==1010=01110=1==0=0=1=10=000=11011001=101===10====11=0110==1==0 
05: Pharaon 3.5.1        48,0/100  ==100000101==1=0010011=0110=10000=010101=0010=1==111===0=10=1=010=000111=11110=001=0=1=10==101100001 
06: SmarThink            47,0/100  =00=00101001100=011110=1=10=10101=11=010=111100=0111=101010=0=001=010001001000===10=011=100=01=0=100 
07: Spike10a             45,5/100  0==00001=0011010=111==1=0=01===101010110=00=1=010=00=00=1===11=0100=0==100=011010=0=1=1=10=0=01=0=01 
08: Colossus2007c        42,5/100  0=0==0=1100011==000=010=111=10001011001100110011010101=1==0=1=00000=0001=0=1011100010100=0000=101010 
09: Pseudo07c            41,5/100  =0100=0=11=1=10=1101===00001=101=00=11011==1010000=0====00010000=101=0=0010001==00000=0=1100=1110001 
09: Booot413             41,5/100  01110==010000=001=1101=1100101101==0=1000=00010=000100=0==0001001001=101==10=00=10==10=0=11=10100100 

    Engine               Score                                                                                                       Ha
01: Hamsters081devF      458,5/900 .................................................................................................... 
02: Glaurung 1.2.1       58,5/100  10=10=000=01=11=0=0=01==1=010111110=01101111=0111=111011=1111==0=11=0=0==0111011==1101=0100=011=10=0 
03: AlaricWB704          53,0/100  11111=1010010011110=0=01=1011011011=000==1=101010=1=0000=0010==00=1=0==0=1011===1=0110=01=00111=1=11 
03: SlowChess Blitz WV21 53,0/100  11==10=0==0110110==111=0011=0=111100=1000101010=0111=1=000====1101=1=01100=000=111001=0100==1110011= 
05: Pharaon 3.5.1        48,5/100  =10010001=0101=10=01===011101101=011111110010=0=0101001001=01=0=01=010=111=0001==1==1===01000=000=10 
05: SmarThink            48,5/100  1100=0100==1=00=01=01=0=0011==1111=00011110===00==110101010=010=11010101101=0100011==1=0000=110==010 
07: Booot413             48,0/100  ===10=01=001010111=1110==0010==00=0=010=01==0101=001110=0=0===0=1=0====1==1=00=100=1111=0011=00=101= 
08: Spike10a             47,0/100  0=10=0===111==100011=1==0=0000011==0101110011==10=010100111=101====110=00=00010=0101=01010010011000= 
09: Colossus2007c        45,5/100  10=0=001110=0101001==10=110010==01010010000100=000001011011101==0001011000100111010==101011110=1001= 
10: Pseudo07c            39,5/100  001110=000=0=00===00100=0=0011=0100=10=1==0==111101=00000110=00==00=00=1000000==100====1=0010==111=1 

Some test results, first 900 games are for 0.6 and next two batches of 900 for 0.7.1, all played at 1+1... small difference but 0.7.1 should be better! Latest version got 477.5 in the first run, it's playing the second batch... hope it holds.
Another explanation: Hamsters is a decent blitzer but sucks at long time controls. After all, there's a limit to what you can do with a 1g brain...