diff options
Diffstat (limited to 'CLAUDE.md')
| -rw-r--r-- | CLAUDE.md | 11 |
1 files changed, 7 insertions, 4 deletions
@@ -28,25 +28,28 @@ The toolchain comes from a Nix flake (`flake.nix`, loaded via direnv `.envrc` ↠## Architecture - `src/board.rs` — all game logic and its unit tests (`#[cfg(test)] mod tests` at the bottom). -- `src/player.rs` — the `Player` trait. Players are polled, not blocking: the coordinator calls `update(board, turn)` before each draw, `pick_move(board, turn) -> Option<Move>` once per loop (computer players search and return `Some`), and passes clicks/keys to `handle_input(Input, board, turn) -> Option<Move>`. `highlights()` gives points for the board to highlight, `reset()` is called on both players after a revert (drop any selection and last-move comments; it is *not* called after a normal move, so players clear their own selection when they return a move), `render(area, buf)` draws the current player's panel (default: its name), and `render_last_turn(area, buf)` draws end-of-turn comments on the move it just made (default: nothing). `impl Widget for &dyn Player` makes every player a ratatui widget. Also the private-ish `MinimaxPlayer` trait: implementors supply `evaluate(board, turn)` (positive favours tigers; `turn` is the side to move) and optionally `depth()` (at most `MAX_DEPTH` = 64)/`time_limit()`/`name()`, plus `set_last_search`/`last_search` to keep a `SearchReport` (move, depth reached, score) that the blanket impl's `render_last_turn` shows; a blanket impl makes every `MinimaxPlayer` a `Player`. +- `src/player.rs` — the `Player` trait. Players are polled, not blocking: the coordinator calls `update(board, turn)` before each draw, `pick_move(board, turn, previous) -> Option<Move>` once per loop (computer players search and return `Some`; `previous` is the game's boards before this one, oldest first, for spotting repetitions), and passes clicks/keys to `handle_input(Input, board, turn) -> Option<Move>`. `highlights()` gives points for the board to highlight, `reset()` is called on both players after a revert (drop any selection and last-move comments; it is *not* called after a normal move, so players clear their own selection when they return a move), `render(area, buf)` draws the current player's panel (default: its name), and `render_last_turn(area, buf)` draws end-of-turn comments on the move it just made (default: nothing). `impl Widget for &dyn Player` makes every player a ratatui widget. Also the private-ish `MinimaxPlayer` trait: implementors supply `evaluate(board, turn)` (positive favours tigers; `turn` is the side to move) and optionally `depth()` (at most `MAX_DEPTH` = 64)/`time_limit()`/`name()`, plus `set_last_search`/`last_search` to keep a `SearchReport` (move, depth reached, score) that the blanket impl's `render_last_turn` shows; a blanket impl makes every `MinimaxPlayer` a `Player`. - Search: alpha-beta (`minimax` checks the cache, `minimax_uncached` searches) with iterative deepening in `pick_move` until half the time limit has passed. Moves are tried in this order: the cache's best move (the `hint`), then tiger jumps (`get_moves` lists them first), then killer moves, then the rest. This is done by rotating moves to the front, because a `sort_by_key` here cost more than it saved. + - Repetitions: `minimax` and `minimax_uncached` carry `path`, the game's earlier boards (from `previous`) followed by the search's own, pushed and popped around each child. A non-root position that repeats one on the path (`repetitions(path, &board) > 0`, so twofold, not the game's threefold) scores `DRAW` = 0 before the cache is consulted. The cache doesn't know the path, so an entry influenced by a repetition can be slightly off for the same board reached another way (accepted, as in chess engines). Every path holds the root, so this can't happen below depth 5 and the depth-3/4 tests against `reference` (which applies the same rule) stay exact. - `cache()` returns an optional `ScoreCache`, keyed by board, side to move and remaining depth. `minimax` passes it `board.canonical()`, so the 8 symmetric boards share one entry, and turns the entry's best move between orientations with `transform_move`/`inverse_symmetry`. This assumes `evaluate` gives symmetric boards the same score. `ClaudePlayer`'s does; a `Network` only does so approximately, since it learns symmetry from augmented data. Entries are `CacheEntry` bounds. `TranspositionTable` is the fixed-size one (`new(bits)`: `1 << bits` slots of ~24 bytes, in two-slot buckets where one slot keeps the deepest entry; `clear()` bumps a generation number instead of wiping). The `HashMap` impl is kept as a simple reference for tests. - `killers()` returns optional `Killers`: two quiet moves per remaining depth that last caused a cutoff. `pick_move` clears both the cache and the killers. - Tests compare the search with plain minimax (`reference`) for every cache kind (`CacheKind::None`/`Map`/`TinyTable`, a table small enough to collide constantly) with and without killers. -- `src/game.rs` — `GameHandler::play_game(sheep_player, tiger_player)` runs a ratatui TUI (mouse capture on) until the user quits, returning the winner if there was one. Layout: board | move history on top; below, the current player's panel (70%) | the previous player's last-turn panel (30%). Panes overlap by one cell (`Layout::spacing(-1)`) and are built with `panel()` (`MergeStrategy::Exact`) so shared borders join up. The bottom row is drawn first and keeps its titles on its bottom border, because its top border is the same line as the board's bottom border, which lists the keybinds (keep them in step with `handle_event`). History is `Vec<(Move, GameBoard)>` (board *after* each move); the side to move comes from its length's parity (sheep first). Clicking a history row calls `revert(i)`, which goes back to the board *before* move `i`. The loop only waits for input (`event::read`) when `pick_move` returns `None`. After a computer's search, `pending_action` handles any input that came in meanwhile, so quit and revert still work in computer-vs-computer games. They take effect once the current search finishes. +- `src/game.rs` — `GameHandler::play_game(sheep_player, tiger_player)` runs a ratatui TUI (mouse capture on) until the user quits, returning the `Outcome` if the game ended. `outcome()` is a win from `check_win`, else `Draw` once `repetitions(&previous(), &board()) >= REPETITION_DRAW`. Layout: board | move history on top; below, the current player's panel (70%) | the previous player's last-turn panel (30%). Panes overlap by one cell (`Layout::spacing(-1)`) and are built with `panel()` (`MergeStrategy::Exact`) so shared borders join up. The bottom row is drawn first and keeps its titles on its bottom border, because its top border is the same line as the board's bottom border, which lists the keybinds (keep them in step with `handle_event`). History is `Vec<(Move, GameBoard)>` (board *after* each move); the side to move comes from its length's parity (sheep first). Clicking a history row calls `revert(i)`, which goes back to the board *before* move `i`. The loop only waits for input (`event::read`) when `pick_move` returns `None`. After a computer's search, `pending_action` handles any input that came in meanwhile, so quit and revert still work in computer-vs-computer games. They take effect once the current search finishes. - `src/board-view.rs` (module `board_view`) — `BoardView` widget (board with rank/file labels and highlights) and `pos_at`, which maps a mouse cell to the nearest point. Both pick the scale from the area (`scale_to_fit`: the largest whole-number scale whose `size(scale)` fits), so the coordinator sizes `board_area` to exactly `size(scale)` and the two agree. Pieces grow with the scale (`reach`): `T` at 1, `[T]` at 2, a 3x3 box from 3, which `size`, `cell` and `pos_at` all allow for. The history pane is a fixed 20 columns; the board pane gets the rest. - `src/local-player.rs` (module `local_player`, pulled in via `#[path]` because of the hyphen) — `LocalPlayer`, a human using the mouse: click a point to place a sheep, or click a piece then one of its highlighted destinations. - `src/claude-player.rs` — `ClaudePlayer`, the minimax computer player: `time_limit` and `max_depth` are public fields, with a 2^20-slot `TranspositionTable` and killer moves. - `src/bench.rs` — the `bench` mode (see above). - `src/network.rs` — `Network`, a small MLP (52 inputs → 64 ReLU → 32 ReLU → tanh, positive favours tigers) with hand-written backprop and an `Adam` optimiser, all in one flat `params` Vec at the `W1`/`B1`/… offsets. Inputs are only the raw position: a bit per point for tigers and for sheep, sheep in hand / 20, and whether tigers are to move. `forward` adds up the layer-1 weight rows of the set bits rather than multiplying through the zeros. Also `Rng` (xorshift64*) and `Sample` (board, turn, target). File format (`save`/`load`): `TSNN`, then version, inputs, hidden1, hidden2 as little-endian u32s, then every param as a little-endian f32; changing the layer sizes makes old files fail to load. - `src/neural-player.rs` — `NeuralPlayer`, a `MinimaxPlayer` like `ClaudePlayer` whose `evaluate` is a `Network`. -- `src/train.rs` — self-play training. Each generation plays `games_per_generation` games of the net against itself at a fixed depth (default 3) across all cores. Moves are random for the first `random_plies`, and with chance `epsilon` after that. A game reaching `ply_cap` is a draw, since the rules have no draw. Each position is labelled with a TD(λ) target worked backwards from the result (±1, or 0) through the searched scores. A position where a random move was played is labelled with its own search score instead, which cuts the chain there. Without that cut, the sheep's random moves (which usually hang a sheep) taught the net that tigers nearly always win. Self-play still ends mostly in tiger wins, because of those random moves, so judge progress by the match against `ClaudePlayer` (`match_games` = 100). The net then trains on the last `buffer_generations` generations, each position shown in a random one of the board's 8 symmetries (`GameBoard::transformed`). With `balance_outcomes` (the default), `balanced` builds each epoch from as many positions of sheep-win games as of tiger-win games: their mean, so the epoch stays the same size. Tiger-win positions are subsampled and sheep-win ones repeated; draws are kept as they are. The output shows the mean error on tiger-win (T) and sheep-win (S) positions separately, before and after training. After that it plays matches against the previous generation and against `ClaudePlayer`, which is only a yardstick and never a training signal. Each generation is saved as `out_dir/gen-NNN.bin` and `latest.bin`, and a rerun carries on from the newest `gen-NNN.bin`. `Config::from` (`--from`) starts from any weights file instead. Generation numbers still follow on from the out dir. The Adam state and the sample buffer are not saved, so they start empty either way. -- `src/main.rs` — the command line, parsed with clap's derive API (`Cli`, `Command`, `PlayArgs`, `TrainArgs`). With no command, or `play`, it plays a game in the TUI and prints the winner; `--sheep`/`--tiger` take `human`, `claude` or a weights file (`parse_player` → `PlayerKind`, turned into a `Box<dyn Player>` by `make_player`), and `--time` is each computer player's time limit in seconds. `neural <weights> [sheep|tiger]` is shorthand for you against a `NeuralPlayer`. `bench [depth]` runs the benchmark. `train [generations] [games] [out dir]` trains a network (use `--release`); every other `train::Config` field is a `--long-option` whose default is read from `Config::default()`, so keep `TrainArgs` and its `From` impl in step with `Config`. +- `src/train.rs` — self-play training. Each generation plays `games_per_generation` games of the net against itself at a fixed depth (default 3) across all cores. Moves are random for the first `random_plies`, and with chance `epsilon` after that. A game is a draw on a threefold repetition (`play` keeps the earlier boards and passes them to `choose`, so the search sees them too), or on reaching `ply_cap`, a backstop for long games that never repeat. Each position is labelled with a TD(λ) target worked backwards from the result (±1, or 0) through the searched scores. A position where a random move was played is labelled with its own search score instead, which cuts the chain there. Without that cut, the sheep's random moves (which usually hang a sheep) taught the net that tigers nearly always win. Self-play still ends mostly in tiger wins, because of those random moves, so judge progress by the match against `ClaudePlayer` (`match_games` = 100). The net then trains on the last `buffer_generations` generations, each position shown in a random one of the board's 8 symmetries (`GameBoard::transformed`). With `balance_outcomes` (the default), `balanced` builds each epoch from as many positions of sheep-win games as of tiger-win games: their mean, so the epoch stays the same size. Tiger-win positions are subsampled and sheep-win ones repeated; draws are kept as they are. The output shows the mean error on tiger-win (T) and sheep-win (S) positions separately, before and after training. After that it plays matches against the previous generation and against `ClaudePlayer`, which is only a yardstick and never a training signal. Each generation is saved as `out_dir/gen-NNN.bin` and `latest.bin`, and a rerun carries on from the newest `gen-NNN.bin`. `Config::from` (`--from`) starts from any weights file instead. Generation numbers still follow on from the out dir. The Adam state and the sample buffer are not saved, so they start empty either way. +- `src/main.rs` — the command line, parsed with clap's derive API (`Cli`, `Command`, `PlayArgs`, `TrainArgs`). With no command, or `play`, it plays a game in the TUI and prints the result; `--sheep`/`--tiger` take `human`, `claude` or a weights file (`parse_player` → `PlayerKind`, turned into a `Box<dyn Player>` by `make_player`), and `--time` is each computer player's time limit in seconds. `neural <weights> [sheep|tiger]` is shorthand for you against a `NeuralPlayer`. `bench [depth]` runs the benchmark. `train [generations] [games] [out dir]` trains a network (use `--release`); every other `train::Config` field is a `--long-option` whose default is read from `Config::default()`, so keep `TrainArgs` and its `From` impl in step with `Config`. Notation (`pos_name`, `impl Display for Move` in board.rs): files `a`–`e` are x 0–4, ranks are `5 - y` (y = 0 is drawn at the top as rank 5). Moves are `sc3` (place), `sc3c2` (sheep move), `ta5b4` (tiger step), `ta5xc3` (tiger jump). Win check: `check_win(turn)` is called with the side about to move, and true means the side that *just* moved has won. Tigers win once 5 sheep are dead (`sheep_count() + sheep_left() <= 15`); otherwise the side that just moved wins if the side to move has no legal moves. That check uses `has_moves(turn)`, which tests the same bitmasks as `get_moves` (shared through `sheep_steps`/`tiger_targets`) without building a list. +Draw: threefold repetition, as in chess. `repetitions(previous, board)` in board.rs counts earlier occurrences of `board` with the same side to move (every second entry of `previous` from the end). Placing and killing can't be undone, so it returns 0 while sheep are in hand and stops scanning at the first board with a different sheep count. A position is drawn when it has come up `REPETITION_DRAW` = 2 times before. `Outcome { Tiger, Sheep, Draw }` (board.rs) is shared by the coordinator and training. + ### Board representation (`GameBoard`) The board is two `u32` bitmasks, `tiger_mask` and `sheep_mask`. This encoding is non-obvious: |