Burke Bot
A chess bot that plays like me — built from every game I've played on
chess.com/Burkeley.
Play it: https://aburke225.github.io/burke-bot/
How it works
pipeline/build.py downloads my full chess.com game history (public API,
rapid games vs humans), merges any PGN files in manual-pgn/ (vs-computer
games — the API never includes those, so they're exported by hand), and
builds docs/book.json: every position I've faced, with the moves I played
there and how often.- While the game is inside that book, the bot samples its move from my real
frequencies — if I play 1. e4 87% of the time, so does the bot.
- Once the game leaves my games, Stockfish 18 (the lite single-threaded WASM
build, vendored) proposes its top 10 moves and a style model picks among
them: a conditional logit trained on my games via behavior cloning
(
pipeline/train_style.py). Ten candidates - not five - so genuinely bad,
human moves are in the pool, and the variable features include
mistake-shaped ones (hangs the piece, leaves one hanging, loses the
exchange) so the model can learn my blunder tendencies, not just my taste.
The trainer analyses with that same vendored engine, run headless through
node — training and play always agree on how candidate moves rank, and
upgrading the vendored engine upgrades both. pipeline/audit.py compares
my mistake profile with the bot's across self-play games.
Every candidate move scores 30 features — 15 static ones that are always
on, and 15 variable ones where the trainer tries all 32,768 subsets and
keeps whichever predicts my moves best on held-out games (with a switching
margin, so near-ties don't churn the model). docs/style.json ships the
winner. Candidates come from a depth-5 search — a human at my level thinks
a couple of moves ahead, so the bot does too. At play time a salience guard
(mirrored in audit.py) drops candidates the aggregate model can't judge:
moves that ignore the threat the opponent just made, voluntary king-walks,
purposeless edge-pawn pushes (queens still on, and the push kicks nothing
off the square it now attacks, defends nothing, is no passer and no storm
at their king — measured: I kick a bishop that is already there 17% of the
time, but push prophylactically to stop one arriving 1.8%, against a 0.7%
baseline for admittedly aimless pushes, and I air out my own castled king
0.4% of the time, which is below that baseline), and — while the position is still
healthy — outright howlers (2.5+ pawns worse than the engine's best;
once already lost, the tail stays, because that is when I flail too).
Each survives only as the engine's #1. And when a near-best capture of a
queen by a lesser piece exists, the bot takes it: that is the one thing
I see every time. If the model fails to load, a strength-capped engine
(skill 4) takes over instead.
- Games played against the bot on the site (only mine, flagged by sign-in)
are captured and folded back in.
- A daily GitHub Action checks for new games and, only when it finds any,
re-runs the pipeline and retrains the model.
Site
docs/ is the whole site — static files on GitHub Pages, no server, no build
step. Cache-busted assets (style.css?v=N, app.js?v=N): bump N on every change.
Rebuilding by hand
python3 -m pip install python-chess numpy
python3 pipeline/build.py
python3 pipeline/train_style.py # analyses with the vendored engine via node
Vendored libraries
