othello/src/ai.rs

422 lines
13 KiB
Rust

use crate::{
board::{Board, explode_board, squares::*},
game::{Game, Team},
table::{Bound, TTEntry, TTable},
};
/// Contains all corner squares
const CORNERS: Board = A1 | A8 | H1 | H8;
/// Contains all edge squares
const EDGES: Board = A2
| A3
| A4
| A5
| A6
| A7
| B1
| B8
| C1
| C8
| D1
| D8
| E1
| E8
| F1
| F8
| G1
| G8
| H2
| H3
| H4
| H5
| H6
| H7;
#[derive(PartialEq, Eq, PartialOrd, Ord)]
/// Represents the _value_ of a move. Some moves at face value
/// better than others.
enum MoveRank {
Corner(Board),
Edge(Board),
Other(Board),
}
impl From<Board> for MoveRank {
fn from(value: Board) -> Self {
// Do bitwise operations to check if we have a
// corner or edge move.
if value & CORNERS > 0 {
Self::Corner(value)
} else if value & EDGES > 0 {
Self::Edge(value)
} else {
Self::Other(value)
}
}
}
impl MoveRank {
/// Unwrap underlying move value out of rank structure
fn into_inner(self) -> Board {
match self {
Self::Corner(m) => m,
Self::Edge(m) => m,
Self::Other(m) => m,
}
}
}
/// Using alpha-beta pruning and the minimax algorithm, determine the best move
/// for a game with a recursion depth of `depth`.
///
/// We use a very simple evaluation heuristic: (Black squares - White squares).
pub fn alphabeta(
mut game: Game,
depth: u8,
mut alpha: i8,
mut beta: i8,
tt: &mut TTable,
) -> (Board, i8, u64) {
let mut num_moves = 0;
// if we reach our maximum recursion depth, return evaluation
if depth == 0 {
return (0, game.score().diff(), num_moves);
}
let moves = game.available();
if moves == 0 {
// if no move, skip and continue recursion
// this seems to technically introduce a bias against move-chains
// that include skips. I haven't found it to be a big deal in play.
game.skip();
return (0, alphabeta(game, depth - 1, alpha, beta, tt).1, num_moves);
}
// just initially assume that the best move is no move at all. This will
// inevitably be corrected.
let mut best_move: Board = 0;
// we initially rank moves based on a couple basic heuristics:
// - corner pieces are best
// - edge pieces are great
// - others considered last
// This just allows us to prune the tree a bit more aggressively
// since we're considering the "best" moves first.
// We do this by mapping moves to ranked moves and then sorting.
let mut moves = explode_board(moves).map(MoveRank::from).collect::<Vec<_>>();
moves.sort_unstable();
let mut moves = moves
.into_iter()
.map(MoveRank::into_inner)
.collect::<Vec<_>>();
// copy our existing alpha/beta for the sake of classifying bounds
let original_alpha = alpha;
let original_beta = beta;
// the brilliance here is that even if we don't have a perfect value
// computed already, the imperfect values still help us get to better values
// quicker.
match tt.get(game.hash) {
Some(entry) if entry.depth >= depth => {
match entry.bound {
// if we know this is exact, trust it without question
Bound::Exact => return (entry.best_move, entry.evaluation, num_moves),
// if we have lower or upper bounds that are more precise than
// our existing alpha and beta values, accept the ones found in
// the cache.
Bound::Lower => alpha = alpha.max(entry.evaluation),
Bound::Upper => beta = beta.min(entry.evaluation),
}
// if we have collapsed the window between alpha and beta, just
// accept the cached entry.
if alpha >= beta {
return (entry.best_move, entry.evaluation, num_moves);
}
// otherwise, if our best move is available, move it to the front
if let Some(best_move_idx) = moves.iter().position(|m| *m == entry.best_move) {
moves[..=best_move_idx].rotate_right(1);
}
}
Some(entry) => {
// otherwise, if our best move is available, move it to the front
if let Some(best_move_idx) = moves.iter().position(|m| *m == entry.best_move) {
moves[..=best_move_idx].rotate_right(1);
}
}
None => {}
}
num_moves = moves.len() as u64;
// I just establish a convention of maximizing for black and minimizing for white.
// I'm not sure if that's conventional or not, but it's what I chose.
match game.current_team {
Team::Black => {
for mv in moves {
let mut g = game.clone();
g.play(mv);
// maximize for the evaluation of subsequent moves
let (_, evaluation, num_moves_sub) = alphabeta(g, depth - 1, alpha, beta, tt);
num_moves += num_moves_sub;
// if our evaluated move is superior to the alpha, update
// it.
if evaluation > alpha {
alpha = evaluation;
best_move = mv;
};
// if our beta is less than alpha, prune the node.
if beta <= alpha {
break;
}
}
let bound = if alpha >= beta {
Bound::Lower
} else if alpha <= original_alpha {
Bound::Upper
} else {
// i.e. alpha < beta || alpha < original_alpha
Bound::Exact
};
tt.store(TTEntry {
depth,
evaluation: alpha,
hash: game.hash,
bound,
best_move,
});
(best_move, alpha, num_moves)
}
Team::White => {
for mv in moves {
let mut g = game.clone();
g.play(mv);
// minimize for the evaluation of subsequent moves
let (_, evaluation, num_moves_sub) = alphabeta(g, depth - 1, alpha, beta, tt);
num_moves += num_moves_sub;
// if our evaluated move produces lower eval than the beta,
// update beta.
if evaluation < beta {
beta = evaluation;
best_move = mv;
};
// if our beta is less than alpha, prune the node.
if beta <= alpha {
break;
}
}
let bound = if beta <= alpha {
Bound::Upper
} else if beta >= original_beta {
Bound::Lower
} else {
Bound::Exact
};
tt.store(TTEntry {
depth,
evaluation: beta,
hash: game.hash,
bound,
best_move,
});
(best_move, beta, num_moves)
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::board::{BitBoard, Board, Score};
use crate::game::{Game, Team};
use rand::prelude::IndexedRandom;
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
/// An AI player that only makes random moves
fn random_move(game: &Game, rng: &mut impl Rng) -> Board {
let moves = explode_board(game.available()).collect::<Vec<_>>();
*moves.choose(rng).unwrap()
}
fn game_after(moves: &[Board]) -> Game {
let mut game = Game::default();
for &mv in moves {
game.safe_play(mv).expect("Move should be valid");
}
game
}
fn assert_ai_move_is_legal(game: &Game, depth: u8) -> Board {
let available = game.available();
let mut tt = TTable::with_mb(2);
let best_move = alphabeta(game.clone(), depth, i8::MIN + 1, i8::MAX - 1, &mut tt).0;
assert_ne!(best_move, 0, "AI should return a move when one exists");
assert_eq!(
best_move & available,
best_move,
"AI returned an illegal move"
);
best_move
}
#[test]
// just a sanity check to ensure that my AI performs up to snuff with another popular engine
fn opening() {
let mut game = Game::default();
let mut tt = TTable::with_mb(24);
game.play(D3);
let (best_move, _, _) = alphabeta(game.clone(), 14, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert_eq!(best_move, C3);
}
#[test]
fn ai_returns_legal_moves_across_curated_positions() {
let cases = vec![
(game_after(&[]), 4),
(game_after(&[D3]), 4),
(game_after(&[F5, F6, E6, F4]), 4),
];
for (game, depth) in cases {
let available = game.available();
if available == 0 {
continue;
}
let mv = assert_ai_move_is_legal(&game, depth);
assert_ne!(mv, 0);
}
}
#[test]
fn ai_prefers_forced_corner() {
let board = BitBoard::from_jon("5bw//////").expect("Valid board");
let game = Game::from_parts(Team::Black, board);
assert_eq!(game.available(), H8);
let mv = assert_ai_move_is_legal(&game, 3);
assert_eq!(mv, H8);
}
#[test]
fn ai_passes_when_no_moves_exist() {
let board = BitBoard::from_jon("wwwwwwww/wwwwwwww/////").expect("Valid board");
let mut tt = TTable::with_mb(2);
let game = Game::from_parts(Team::Black, board);
assert_eq!(game.available(), 0);
let (mv, eval, _) = alphabeta(game.clone(), 4, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert_eq!(mv, 0);
assert_eq!(eval, game.score().diff());
}
#[test]
fn tt_exact_root_hit_eliminates_repeat_search() {
let game = Game::default();
let mut tt = TTable::with_mb(2);
let (best_move, eval, first_considered) =
alphabeta(game.clone(), 1, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert!(first_considered > 0);
let (cached_move, cached_eval, second_considered) =
alphabeta(game.clone(), 1, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert_eq!(cached_move, best_move);
assert_eq!(cached_eval, eval);
assert_eq!(second_considered, 0);
}
#[test]
fn tt_lower_bound_hit_still_searches_with_wide_window() {
let game = Game::default();
let mut tt = TTable::with_mb(2);
tt.store(TTEntry {
bound: Bound::Lower,
evaluation: 0,
depth: 1,
best_move: D3,
hash: game.hash,
});
let (_, _, considered) = alphabeta(game.clone(), 1, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert!(considered > 0);
}
#[test]
fn tt_upper_bound_hit_still_searches_with_wide_window() {
let game = Game::default();
let mut tt = TTable::with_mb(2);
tt.store(TTEntry {
bound: Bound::Upper,
evaluation: 0,
depth: 1,
best_move: D3,
hash: game.hash,
});
let (_, _, considered) = alphabeta(game.clone(), 1, i8::MIN + 1, i8::MAX - 1, &mut tt);
assert!(considered > 0);
}
// I found that, despite the AI clobbering me, the AI could not
// compete with itself very well. I'm honestly not quite sure why that is.
#[test]
fn ai_beats_random() {
// just contains pairings of starting_team and seed value
let cases = vec![
(Team::Black, 1231293),
(Team::White, 491823),
(Team::White, 12931),
(Team::Black, 982983713),
(Team::Black, 123),
(Team::White, 87132895),
];
let mut tt = TTable::with_mb(2);
for (team, seed) in cases {
let mut rng = StdRng::seed_from_u64(seed);
let mut game = Game::default();
if team != Team::Black {
let mv = random_move(&game, &mut rng);
game.play(mv);
}
while !game.is_complete() {
if game.available() == 0 {
game.skip();
continue;
}
let mv = if game.current_team == team {
alphabeta(game.clone(), 8, i8::MIN + 1, i8::MAX - 1, &mut tt).0
} else {
random_move(&game, &mut rng)
};
assert_eq!(mv & game.available(), mv, "AI generated an illegal move");
game.play(mv);
}
assert!(
match (team, game.score()) {
(Team::Black, Score(b, w)) => b - w,
(Team::White, Score(b, w)) => w - b,
} > 4,
"game with seed {} and team {:?} failed to win by 4 points or more.",
seed,
team
);
}
}
#[test]
fn move_ordering() {
let mv = A1 | A8 | C3 | D5 | A4;
let mut moves = explode_board(mv).map(MoveRank::from).collect::<Vec<_>>();
moves.sort();
let moves = moves
.into_iter()
.map(MoveRank::into_inner)
.collect::<Vec<_>>();
assert_eq!(moves, vec![A1, A8, A4, C3, D5]);
}
}