The following three code samples demonstrate some of the flexibility of the Ceres API for conducting chess research efficiently. The code snippets were written inside the very convenient Linqpad tool (https://www.linqpad.net/):
This first simple sample dumps an overview of the single node analysis of the position after Bobby Fischer's infamous blunder in Game 1 of the World Championship match. It demonstrates that progressively more advanced versions of the Leela Chess Zero networks do a monotonically increasing better job of understanding in a single-node analysis that the pawn move g2g3 trapping the bishop is best.
703810: 10x128 WDL MLH <NNPositionEvaluation V= -0.18 Policy= g2g3 [ 27.26%] e3e4 [ 25.91%] f2f4 [ 12.23%] c1b2 [ 10.50%] a3a4 [ 7.08%] >
32930: 20x256 <NNPositionEvaluation V= -0.10 Policy= g2g3 [ 57.32%] f2f4 [ 26.45%] a3a4 [ 4.70%] e3e4 [ 4.22%] c1d2 [ 1.35%] >
42767: 20x256 <NNPositionEvaluation V= -0.16 Policy= g2g3 [ 75.53%] f2f4 [ 6.50%] e3e4 [ 5.73%] a3a4 [ 3.82%] d3e2 [ 2.82%] >
66511: 24x320 WDL MLH <NNPositionEvaluation V= 0.01 Policy= g2g3 [ 82.32%] a3a4 [ 3.68%] e3e4 [ 3.56%] f2f4 [ 3.19%] c1b2 [ 1.37%] >
66666: 30x384 WDL MLH <NNPositionEvaluation V= 0.15 Policy= g2g3 [ 85.61%] f2f4 [ 7.91%] e3e4 [ 1.30%] a3a4 [ 1.07%] d3e2 [ 0.72%] >
66846: 30x384 WDL MLH <NNPositionEvaluation V= 0.20 Policy= g2g3 [ 87.23%] f2f4 [ 6.93%] e3e4 [ 1.05%] a3a4 [ 0.88%] d3e2 [ 0.66%] >
J94-80: 30x384 WDL MLH <NNPositionEvaluation V= 0.51 Policy= g2g3 [ 90.22%] f2f4 [ 4.09%] a3a4 [ 0.98%] e3e4 [ 0.98%] d3e2 [ 0.70%] >
The second sample extracts a large subset of games from a set of top chess players in history and compares their actual moves made against an LC0 policy network, quantifying the agreement to assess their strength and/or similarity to the playing style of the Leela networks.
After a couple of minutes of execution the following intriguing results are output:
60.91% Anand in 26201 moves
59.75% Fischer in 6760 moves
59.74% Carlsen in 36374 moves
59.62% Kasparov in 24645 moves
58.89% Capablanca in 5519 moves
58.29% Tal in 15604 moves
58.20% Caruana in 56363 moves
57.78% Topalov in 24941 moves
57.71% Euwe in 7543 moves
57.50% Kramnik in 31972 moves
56.47% Korchnoi in 17275 moves
56.42% Lasker in 9844 moves
56.39% Nakamura in 66737 moves
56.15% Karpov in 38493 moves
55.39% Smyslov in 13718 moves
55.22% Portisch in 15675 moves
54.95% Huebner in 12144 moves
54.94% Alekhine in 21984 moves
54.85% Botvinnik in 5404 moves
54.77% Reshevsky in 8059 moves
53.66% Pachman in 4210 moves
52.07% Petrosian in 13968 moves
The third example analyzes one chess position using multiple engines including Stockfish, Leela, the default Ceres engine, and a version of the Ceres engine configured with a different set of CPUCT tuning parameters. Detailed root move statistics are dumped for each of the neural network based searches to facilitate analysis and interpretation.
POSITION EVALUATION USING Stockfish 12: info depth 36 seldepth 58 multipv 1 score cp 443 nodes 43750282 nps 1662055 hashfull 1000 tbhits 0 time 26323
pv f3e5 b6d6 b5b6 f5e4 f4d2 f7e5 d2g5 e5g6 c1c7 f8f7 c7c8 f7f8 g5h6 d6e7 f6e7 f8c8 e7g5 c8a8 f2f3 e4f5
POSITION EVALUATION USING Leela Chess Zero:
<LC0VerboseMoveStats node Visits=50226 V = 0.4464 P=100 Q=0, U=0 [48]
<LC0VerboseMoveStats f3e5 Visits=45309 V = 0.7364 P=16.68 Q=0, U=0.00484 [588]
<LC0VerboseMoveStats f6e7 Visits=1400 V = 0.4465 P=8.93 Q=0, U=0.08379 [1317]
<LC0VerboseMoveStats c1c7 Visits=1255 V = 0.6349 P=14.45 Q=0, U=0.15133 [70]
<LC0VerboseMoveStats f6d4 Visits=876 V = 0.6955 P=9.85 Q=0, U=0.14777 [1299]
<LC0VerboseMoveStats f6b2 Visits=498 V = 0.5347 P=3.87 Q=0, U=0.10186 [1295]
<LC0VerboseMoveStats f6a1 Visits=166 V = 0.4423 P=2.71 Q=0, U=0.21317 [1293]
...
POSITION EVALUATION USING <NNEvaluatorDef Network=LC0:J96-44 Device=GPU:0 >
<NNPositionEvaluation V= 0.45 Policy= f3e5 [ 28.62%] c1c7 [ 22.64%] f6d4 [ 12.33%] f6e7 [ 10.52%] f6c3 [ 5.92%] >
info depth 12 seldepth 23 time 9639 nodes 41858 score cp 63 tbhits 0 nps 4343 pv f3e5 b6d6 b5b6 f8a8 g3g4 h5g4 c1c7 d3d2 c7f7 d2d1q g1h2
VERBOSE ROOT MOVE STATISTICS
0 (none) 100.00% N= 41858 Q=-0.389 V= 0.446 WDL= 0.53 0.39 0.08 WDL Avg= 0.53 0.32 0.14 M = 78 63
1 f3e5 16.89% N= 34811 Q= 0.428 V= 0.730 WDL= 0.04 0.20 0.77 WDL Avg= 0.14 0.30 0.56 M = 60 58
S 1 c1c7 14.59% N= 2086 Q= 0.256 V= 0.632 WDL= 0.05 0.27 0.68 WDL Avg= 0.18 0.39 0.43 M = 66 68
S 1 f6d4 10.02% N= 1408 Q= 0.256 V= 0.692 WDL= 0.04 0.23 0.73 WDL Avg= 0.17 0.41 0.42 M = 70 89
S 1 f6e7 9.07% N= 1350 Q= 0.258 V= 0.446 WDL= 0.08 0.40 0.52 WDL Avg= 0.12 0.50 0.38 M = 78 94
S 1 f6b2 3.92% N= 547 Q= 0.256 V= 0.534 WDL= 0.06 0.35 0.59 WDL Avg= 0.14 0.47 0.40 M = 76 95
S 1 f3g5 5.79% N= 353 Q= 0.160 V= 0.494 WDL= 0.07 0.36 0.57 WDL Avg= 0.17 0.50 0.33 M = 74 88
S 1 f6c3 6.35% N= 351 Q= 0.146 V= 0.483 WDL= 0.07 0.39 0.55 WDL Avg= 0.14 0.57 0.29 M = 81 90
S 1 f6a1 2.75% N= 210 Q= 0.196 V= 0.442 WDL= 0.07 0.42 0.51 WDL Avg= 0.15 0.51 0.34 M = 81 92
S 1 g3g4 2.94% N= 113 Q= 0.069 V= 0.538 WDL= 0.06 0.34 0.60 WDL Avg= 0.23 0.47 0.30 M = 66 76
S 1 f3d4 3.55% N= 102 Q=-0.027 V= 0.155 WDL= 0.15 0.55 0.30 WDL Avg= 0.25 0.53 0.22 M = 89 75
(followed by 38 additional moves not shown...)
-----------------------------------------------------------------------------------------------------------------
POSITION EVALUATION USING ZZTUNE2 of <NNEvaluatorDef Network=LC0:J96-44 Device=GPU:0 >
<NNPositionEvaluation V= 0.45 Policy= f3e5 [ 28.62%] c1c7 [ 22.64%] f6d4 [ 12.33%] f6e7 [ 10.52%] f6c3 [ 5.92%] >
info depth 7 seldepth 15 time 7421 nodes 46245 score cp 25 tbhits 0 nps 6231 pv f6e7 f8a8 c1c7 d3d2 f3d2 b6b5 e7f6 b5e2 c7b7 a8c8 d2f3
VERBOSE ROOT MOVE STATISTICS
0 (none) 100.00% N= 46245 Q=-0.176 V= 0.446 WDL= 0.53 0.39 0.08 WDL Avg= 0.44 0.30 0.26 M = 78 72
1 f6e7 9.07% N= 31844 Q= 0.226 V= 0.451 WDL= 0.08 0.40 0.53 WDL Avg= 0.23 0.30 0.46 M = 78 72
S 1 f3e5 16.89% N= 5062 Q= 0.095 V= 0.730 WDL= 0.04 0.20 0.77 WDL Avg= 0.30 0.30 0.40 M = 60 69
S 1 f6d4 10.02% N= 2595 Q= 0.095 V= 0.692 WDL= 0.04 0.23 0.73 WDL Avg= 0.31 0.28 0.41 M = 70 74
S 1 c1c7 14.59% N= 1751 Q= 0.019 V= 0.632 WDL= 0.05 0.27 0.68 WDL Avg= 0.39 0.21 0.41 M = 66 44
S 1 f6c3 6.35% N= 967 Q= 0.089 V= 0.483 WDL= 0.07 0.39 0.55 WDL Avg= 0.28 0.36 0.37 M = 81 72
S 1 f6b2 3.92% N= 802 Q= 0.099 V= 0.535 WDL= 0.06 0.35 0.59 WDL Avg= 0.28 0.33 0.38 M = 76 74
S 1 f6e5 1.47% N= 678 Q= 0.109 V= 0.335 WDL= 0.09 0.49 0.42 WDL Avg= 0.27 0.35 0.38 M = 73 67
S 1 f6a1 2.75% N= 563 Q= 0.098 V= 0.442 WDL= 0.07 0.42 0.51 WDL Avg= 0.29 0.32 0.39 M = 81 74
S 1 f3g5 5.79% N= 305 Q=-0.076 V= 0.494 WDL= 0.07 0.36 0.57 WDL Avg= 0.43 0.23 0.35 M = 74 63
S 1 f3d4 3.55% N= 208 Q=-0.016 V= 0.155 WDL= 0.15 0.55 0.30 WDL Avg= 0.32 0.37 0.31 M = 89 70
(followed by 38 additional moves not shown...)
-----------------------------------------------------------------------------------------------------------------