Skip to content

tasks.md

Latest commit

 

History

History
97 lines (73 loc) · 4.48 KB

tasks.md

File metadata and controls

97 lines (73 loc) · 4.48 KB
Raw

Tasks

The next five changes, roughly in priority order (Elo per unit of effort, and how much each compounds with the rest). Each is a self-contained addition to Negamax / lazy_smp / the evaluator unless noted. Elo figures are rough community numbers for an engine already at this level - measure, don't trust.

1. Aspiration windows

Iterative deepening currently re-searches every depth with the full (-INF, INF) window. Instead, open depth d with a narrow window around the previous iteration's score (score ± 25, say); on a fail-high or fail-low widen that side (double the delta, or jump to ±INF) and re-search. Most iterations land inside the window and search far fewer nodes. Lives in lazy_smp._worker's deepening loop, per worker. Pairs naturally with PVS. ~15-30 Elo, a small change.

2. Shallow-depth pruning (reverse futility, futility, late move pruning)

At low depth, near the leaves, prune aggressively on the static eval:

  • Reverse futility / static null move: if eval - margin*depth >= beta at depth <= ~6 and not in check, return eval.
  • Futility: at depth <= ~2, if eval + margin < alpha, skip quiet moves that can't raise alpha.
  • Late move pruning: past a depth-dependent move count at low depth, skip the remaining quiet non-checking moves entirely.

Never when in check, for captures/promotions, or near mate scores. Big node reduction; the effective branching factor drops again. ~40-70 Elo combined, but tune the margins carefully - too greedy and tactics start getting missed (guard with the existing mate tests plus a WAC/ECM tactical suite).

3. Static Exchange Evaluation (SEE)

A see(board, move) -> int that plays out the capture sequence on one square with the cheapest attacker each time. Two uses:

  • Move ordering: order captures with SEE < 0 after the quiet killers instead of just behind winning captures - MoveOrderer currently trusts MVV-LVA, which mis-ranks a queen grabbing a defended pawn.
  • Quiescence pruning: in Negamax.quiesce, skip captures with SEE < 0 entirely instead of searching every capture.

python-chess gives board.attackers(color, square) to build the attacker lists. ~25-50 Elo, and it makes quiescence much cheaper.

4. Search extensions

Spend an extra ply where the tree is forcing so tactics are not missed at the horizon:

  • Check extension: depth += 1 when the move gives check (cap total extensions per line so it can't blow up).
  • One-reply extension: extend when the side to move has a single legal move.
  • Later: singular extensions (re-search to prove the TT move is the only good one) - higher effort, do it after PVS/LMR are stable.

~15-25 Elo for the cheap two; more with singular.

5. Evaluation tuning harness (Texel's method)

The PeSTO tables and the new eval_terms weights are untuned for this engine. Add tools/tune.py: label a few hundred thousand quiet positions with the game result (from self-play PGNs or a public dataset), then fit every weight by minimising the logistic error between sigmoid(eval) and the result. Keep the weights in one place so the tuner can rewrite them. Unlocks the eval terms already added and makes every future term measurable. ~30-100 Elo depending on how untuned things currently are; the highest-ceiling item on this list.

Also worth doing

  • Syzygy tablebases (chess.syzygy): probe WDL/DTZ for <= 6 pieces at the root and in the search for perfect endgame play. Needs a SyzygyPath and the tablebase files (~150 GB for 6-man), so gate it on config and no-op when the files are absent.
  • UCI setoption: Hash, Threads, Contempt, SyzygyPath, MultiPV - currently parsed and ignored.
  • 2-fold repetition as a draw inside the search (not just 3-fold): a position seen twice within the tree is almost always a forced draw; detecting it a repetition earlier saves nodes.

Housekeeping

GitHub repo settings (do these in the repo UI - not tracked in files)

  • Rename the repo pyChess -> pychess so it matches the package name; update the local remote afterwards.

Opening book (opening_book/bookfish.bin)

An ~18 MB binary lives in git history (see opening_book/README.md). It works fine as-is, but for a cleaner repo consider one of: migrate it to Git LFS, fetch it on first run instead of vendoring it, or document it as an optional external download. Any of these needs a history rewrite to actually shrink the pack, so it is a deliberate call, not a drive-by fix.