Chess Fighter 2.0 release

[Alexander Lim]

Dear all,

Chess Fighter 2.0 is now available. Here is a link to a zip file containing all the files:

https://drive.google.com/open?id=1i9CTk ... 7xrfCeeynq

I've managed to implement most of what you expect from a UCI engine.
Time controls: blitz, fixed time, infinite analysis, fixed nodes. Does not support fixed depth as does not make sense for MCTS. Will output the PV and info regarding the policy priors.

There is no table base support.

You will need a GPU with at least Cuda 9 installed.

The neural net sizes are 8x128. Training is a mixture of Chess Fighter self-play games and supervised learning from CCRL games. Training is still ongoing.

Chess Fighter does not use history planes so there is no issue with giving it FENS and using it for analysis.

On my 1080ti I'm getting around 7000nps and would estimate its strength around 2800-2900 ELO. It's quite an aggressive player and doesn't shy away from complicated positions despite tactics being its weakest side. Although it can miss certain types of tactics that are easy for classical engines (in particular x-ray attacks), it often finds certain tactical / positional sacrifices that the classical engines just can't find.

If anyone does download, could you please post if Chess Fighter worked on your system and what nps you are getting and your PC/GPU specs. Thanks.

[ChickenLogic]

Works fine under Windows 10 with a 980 TI in cutechess with 4.5 up to 5.5kn/s but not under Arena. Arena's debug endlessly shows ". \n . \n . \n" (\n is a new line).

[dkappe]

Dear all,

Chess Fighter 2.0 is now available. Here is a link to a zip file containing all the files:

https://drive.google.com/open?id=1i9CTk ... 7xrfCeeynq

I've managed to implement most of what you expect from a UCI engine.
Time controls: blitz, fixed time, infinite analysis, fixed nodes. Does not support fixed depth as does not make sense for MCTS. Will output the PV and info regarding the policy priors.

There is no table base support.

You will need a GPU with at least Cuda 9 installed.

The neural net sizes are 8x128. Training is a mixture of Chess Fighter self-play games and supervised learning from CCRL games. Training is still ongoing.

Chess Fighter does not use history planes so there is no issue with giving it FENS and using it for analysis.

On my 1080ti I'm getting around 7000nps and would estimate its strength around 2800-2900 ELO. It's quite an aggressive player and doesn't shy away from complicated positions despite tactics being its weakest side. Although it can miss certain types of tactics that are easy for classical engines (in particular x-ray attacks), it often finds certain tactical / positional sacrifices that the classical engines just can't find.

If anyone does download, could you please post if Chess Fighter worked on your system and what nps you are getting and your PC/GPU specs. Thanks.

Is the training code available?

[AdminX]

Also working on MX150 using Shreddder GUI

[mar]

Works fine under Windows 10 with a 980 TI in cutechess with 4.5 up to 5.5kn/s but not under Arena. Arena's debug endlessly shows ". \n . \n . \n" (\n is a new line).

That's because Arena thinks it's another engine using winboard protocol. Just change it manually to UCI and it will work fine.

[adnoh]

Works fine on my Intel i5-3570K / GTX1060 using Shredder GUI.

In a quick 10 game blitz (1'+1'') it was 5-5 against Senpai 2.0

Here are the results from 1 minute spent analysis the opening position.

Move overhead = 1100 milliseconds
Batch inference size = 50
FPU = -1
Policy softmax temperature = 0.3
PUCT = 1.5
Policy net = preg_8x128b_1454
Value net = v_8x128c_6388
Engine: Chess Fighter 2 (0 MB)
by Alexander Lim
15.00 0:00 +0.29 1.d4 d5 2.Nf3 Nf6 3.g3 g6 4.Bg2 Bg7 5.O-O O-O 6.b3 c5 7.Bb2 cxd4 8.Nxd4 (1.000)
15.00 0:03 +0.30 1.d4 d5 2.Nf3 Nf6 3.Bf4 g6 4.e3 Bg7 5.c4 O-O 6.Nc3 Nc6 7.h3 a6 8.Rc1 (11.000)
18.00 0:04 +0.29 1.d4 d5 2.Nf3 Nf6 3.g3 g6 4.Bg2 Bg7 5.O-O O-O 6.b3 c5 7.Bb2 cxd4 8.Nxd4 e5 9.Nf3 Nc6 (14.000)
19.00 0:12 +0.25 1.d4 d5 2.Nf3 Nf6 3.g3 g6 4.Bg2 Bg7 5.O-O O-O 6.b3 c5 7.Bb2 cxd4 8.Nxd4 e5 9.Nf3 Nc6 10.Nxe5 (44.000)
20.00 0:13 +0.24 1.d4 d5 2.Nf3 Nf6 3.g3 g6 4.Bg2 Bg7 5.O-O O-O 6.b3 c5 7.Bb2 cxd4 8.Nxd4 e5 9.Nf3 Nc6 10.Nxe5 Nxe5 (49.000)
21.00 0:24 +0.26 1.e4 d5 2.exd5 Qxd5 3.Nc3 Qd8 4.d4 Nf6 5.Nf3 g6 6.Bc4 Bg7 7.O-O O-O 8.Re1 Nc6 9.Bf4 Bg4 10.d5 Na5 11.Be2 (89.000)
22.00 0:46 +0.24 1.e4 e5 2.d4 exd4 3.Qxd4 Nc6 4.Qe3 Nf6 5.Nc3 d6 6.Nf3 Be7 7.Be2 O-O 8.O-O Re8 9.h3 d5 10.Nxd5 Nxd5 11.Qb3 Nb6 (170.000)
f2-f3( -0.029042 / 184 / 0.697470% )
g2-g4( -0.040256 / 327 / 1.490777% )
Nb1-a3( -0.023207 / 414 / 1.396262% )
b2-b4( -0.016378 / 439 / 1.267469% )
Ng1-h3( -0.025938 / 442 / 1.572209% )
f2-f4( -0.023546 / 483 / 1.639967% )
a2-a4( -0.008516 / 488 / 1.141740% )
h2-h3( -0.005476 / 772 / 1.638107% )
h2-h4( -0.010851 / 786 / 1.965726% )
a2-a3( -0.003740 / 893 / 1.789624% )
d2-d3( -0.003212 / 1731 / 3.401992% )
c2-c3( 0.001445 / 1845 / 3.019612% )
b2-b3( -0.006594 / 2686 / 5.916355% )
e2-e3( 0.008162 / 4167 / 4.871481% )
g2-g3( 0.005985 / 9737 / 12.945726% )
Nb1-c3( 0.013187 / 10101 / 8.324109% )
Ng1-f3( 0.011128 / 13137 / 12.694550% )
c2-c4( 0.022082 / 18652 / 4.265577% )
d2-d4( 0.021804 / 78643 / 18.389801% )
e2-e4( 0.022864 / 80417 / 11.571458% )
puct = 1.5 temp = 0.3 fpu = -1 total policy = 99.926%
22.00 1:02 +0.22 1.e4 e5 2.d4 exd4 3.Qxd4 Nc6 4.Qe3 Nf6 5.Nc3 d6 6.Nf3 Be7 7.Be2 O-O 8.O-O Re8 9.h3 d5 10.Nxd5 Nxd5 11.Qb3 Nb6 (226.344)
best move: e2-e4 time: 1:02.344 minnodes: 226.344

[ChickenLogic]

Thank you, I didn't notice that Arena expected winboard.

[George Tsavdaris]

You will need a GPU with at least Cuda 9 installed.

Thanks. Where can i find that? Any concrete link to download them?

OR since Leela got permission to have the appropriate dlls distributed with her, is there any chance that putting these 3 dlls of Leela for CUDA and cudnn in a folder with Chessfighter, then Chessfighter to work?
These 3 "Leela's" dlls are:
cublas64_100.dll
cudart64_100.dll
cudnn64_7.dll

[Alexander Lim]

Also working on MX150 using Shreddder GUI

Thanks! Much appreciated.

[Alexander Lim]

Works fine under Windows 10 with a 980 TI in cutechess with 4.5 up to 5.5kn/s but not under Arena. Arena's debug endlessly shows ". \n . \n . \n" (\n is a new line).

That's because Arena thinks it's another engine using winboard protocol. Just change it manually to UCI and it will work fine.

OK problem solved. Was about to ask for a screen shot as I couldn't recreate it on my Arena. No need to now.