ENGINE RESEARCH · OCTOBER 2026
A learned pawn engine.
The engine combines a compact learned evaluator, deeper offline opening searches, and exact pawn endings. Everything used during play runs in your browser.
The selected network
104 → 128 ReLU → 1 value logit, with 13,569 parameters. Inputs encode 48 squares for each side’s pawns and eight en-passant files. The board is oriented from the player to move’s perspective, preserving tempo information. No neural network decides legality or terminal outcomes.
The model was trained on a 455,116-sample corpus drawn from 30,000 attempted self-play games and 12,000 solved small endings. Training combines final outcomes with search-derived targets. Games, including their mirrored positions, stay within their assigned training or validation split.
What was compared
| Experiment | Finding |
|---|---|
| 16, 32, 64, 128, 256 hidden units | 128 units provided a useful balance of prediction quality and browser search cost. |
| Second hidden layer; explicit pawn feature maps | Lower prediction loss did not consistently produce stronger fixed-time play. |
| Gumbel-inspired search and 6,000 self-play games | Competitive at one tested search budget, but no clear improvement over the selected value/search approach. |
| Randomized search budgets; shorter replay windows | Tested as training-efficiency alternatives; neither displaced the selected model. |
| MTD(f), PVS, aspiration windows, late-move reductions | MTD(f) showed a modest gain in pawn-only play. PVS remains the analysis and mixed-piece search. |
Precomputed knowledge
The endgame table covers every placement with two, three, or four pawns on the six playable ranks. Color/turn symmetry covers both sides to move. Live en-passant rights are resolved through exact legal successors. It stores outcomes and distance to a result, favoring faster wins and longer resistance.
Independent JavaScript solving checked every two-pawn placement and 1,500 sampled three/four-pawn positions. The four-pawn table contains 2,724,120 positions and compresses to about 165 KB. A 1,719-position opening book covers the computer’s first three decisions along its chosen replies, with about 1.84 billion nodes searched offline. Book evaluations remain estimates.
Playing tests
Against the previous deployed version, using fresh matched openings and the same time per move:
| Time per move | Games | Win / draw / loss | Score | 95% paired interval |
|---|---|---|---|---|
| 80 ms | 400 | 271 / 64 / 65 | 75.75% | 72.5–78.9% |
| 800 ms | 40 | 23 / 8 / 9 | 67.5% | 58.75–76.25% |
The opening book supplied only seven moves in the 400-game test and none in the 800 ms test, so the improvement extends beyond prepared opening moves. Tests used the actual JavaScript engine in Node; timing on your device will differ.
The expanded knight-variant check was approximately level: 66 wins, 15 draws, 67 losses, and 12 unresolved games. Small bishop and full-material checks also do not establish a reliable strength gain. There is no claim of improved extra-piece evaluation or a demonstrated ability to beat a particular human.
Scores count a win as 1 and a draw as ½. Tests use matched openings with colors swapped. Confidence intervals resample opening pairs. These are engine comparisons, not a calibrated human rating. Discovery experiments guided selection; the final comparison uses fresh seeds.
Download the full technical report
Queen-capture rules update
Current games allow an immediate capture of the first promoted queen, after which play continues. The agreed on-move queen, pawn-or-queen exhaustion, and stalemate criteria are explained in the game’s Rules tab. Pawn-only outcomes and the trained assets are unchanged. The playing comparisons above predate this correction; they are not new mixed-piece results.
Mixed-piece experiment
A later experiment trained a small correction for positions with pieces, compared it against an optimized expert, and tested the pawn engine as a projected oracle. Read the new experiment and fresh results.
Scope and reproducibility
The neural model and tablebase target pawn-only positions. The original experiment used the classical evaluator for knights, bishops, and rooks. The later mixed-piece update is documented separately; this original pawn model still takes over when the extra pieces are gone. All rule handling is shared with the game interface.
Offline training used eight CPU cores, NumPy, and a native rules implementation checked against the browser’s JavaScript rules. No paid GPU service was used. That experiment’s 69 automated tests passed; the native and JavaScript rules also agree on 5,000 positions and 45,567 legal successors. The source repository includes training programs, seeds, experiment configurations, model checkpoints, benchmark records, and validation tests.