chess_inator/src/search.rs

576 lines
18 KiB
Rust

/*
This file is part of chess_inator.
chess_inator is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3 of the License, or (at your option) any later version.
chess_inator is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with chess_inator. If not, see https://www.gnu.org/licenses/.
Copyright © 2024 dogeystamp <dogeystamp@disroot.org>
*/
//! Game-tree search.
use crate::hash::ZobristTable;
use crate::prelude::*;
use std::cmp::{max, min};
use std::sync::mpsc;
use std::time::{Duration, Instant};
// min can't be represented as positive
const EVAL_WORST: EvalInt = -(EvalInt::MAX);
const EVAL_BEST: EvalInt = EvalInt::MAX;
#[cfg(test)]
mod test_eval_int {
use super::*;
#[test]
fn test_eval_worst_best_symm() {
// int limits will bite you if you don't test this
assert_eq!(EVAL_WORST, -EVAL_BEST);
assert_eq!(-EVAL_WORST, EVAL_BEST);
}
}
/// Eval in the context of search.
#[derive(PartialEq, Eq, Clone, Copy, Debug)]
pub enum SearchEval {
/// Mate in |n| - 1 half moves, negative for own mate.
Checkmate(i8),
/// Centipawn score (exact).
Exact(EvalInt),
/// Centipawn score (lower bound).
Lower(EvalInt),
/// Centipawn score (upper bound).
Upper(EvalInt),
/// Search was hard-stopped.
Stopped,
}
impl SearchEval {
/// Flip side, and increment the "mate in n" counter.
fn increment(self) -> Self {
match self {
SearchEval::Checkmate(n) => {
debug_assert_ne!(n, 0);
if n < 0 {
Self::Checkmate(-(n - 1))
} else {
Self::Checkmate(-(n + 1))
}
}
SearchEval::Exact(eval) => Self::Exact(-eval),
SearchEval::Lower(eval) => Self::Upper(-eval),
SearchEval::Upper(eval) => Self::Lower(-eval),
SearchEval::Stopped => SearchEval::Stopped,
}
}
}
impl From<SearchEval> for EvalInt {
fn from(value: SearchEval) -> Self {
match value {
SearchEval::Checkmate(n) => {
debug_assert_ne!(n, 0);
if n < 0 {
EVAL_WORST - EvalInt::from(n)
} else {
EVAL_BEST - EvalInt::from(n)
}
}
SearchEval::Exact(eval) => eval,
SearchEval::Lower(eval) => eval,
SearchEval::Upper(eval) => eval,
SearchEval::Stopped => panic!("Attempted to evaluate a halted search"),
}
}
}
impl Ord for SearchEval {
fn cmp(&self, other: &Self) -> std::cmp::Ordering {
let e1 = EvalInt::from(*self);
let e2 = EvalInt::from(*other);
e1.cmp(&e2)
}
}
impl PartialOrd for SearchEval {
fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
Some(self.cmp(other))
}
}
/// Configuration for the gametree search.
#[derive(Clone, Copy, Debug)]
pub struct SearchConfig {
/// Enable alpha-beta pruning.
pub alpha_beta_on: bool,
/// Limit regular search depth
pub depth: usize,
/// Limit quiescence search depth
pub qdepth: usize,
/// Parameter (centipawns) that sets how confident the engine is.
///
/// Positive means avoid draws, and try to win instead.
///
/// Depending on the game phase, an extra factor will be multiplied too; in the beginning of
/// the game the opponent is more likely to blunder later and lose their advantage, so we don't
/// go for draws. Later, the result is more certain, so reduce the contempt factor.
///
/// An alternative interpretation of this: the contempt factor is the negative of the value
/// assigned to a draw.
pub contempt: EvalInt,
/// Enable transposition table.
pub enable_trans_table: bool,
/// Transposition table size (2^n where this is n)
pub transposition_size: usize,
/// Print machine-readable information about the position during NNUE training data generation.
pub nnue_train_info: bool,
}
impl Default for SearchConfig {
fn default() -> Self {
SearchConfig {
alpha_beta_on: true,
depth: 16,
qdepth: 6,
contempt: 0,
enable_trans_table: true,
transposition_size: 24,
nnue_train_info: false,
}
}
}
/// Least valuable victim, most valuable attacker heuristic for captures.
fn lvv_mva_eval(src_pc: Piece, cap_pc: Piece) -> EvalInt {
let pc_values = [500, 300, 300, 20000, 900, 100];
pc_values[cap_pc as usize] - pc_values[src_pc as usize]
}
/// Assign a priority to a move based on how promising it is.
fn move_priority(board: &mut Board, mv: &Move, state: &mut EngineState) -> EvalInt {
// move eval
let mut eval: EvalInt = 0;
let src_pc = board.get_piece(mv.src).unwrap();
let anti_mv = mv.make(board);
if state.config.enable_trans_table {
if let Some(entry) = &state.cache[board.zobrist] {
eval = entry.eval.into();
}
} else if let Some(cap_pc) = anti_mv.cap {
// least valuable victim, most valuable attacker
eval += lvv_mva_eval(src_pc.into(), cap_pc)
}
anti_mv.unmake(board);
eval
}
/// State specifically for a minmax call.
struct MinmaxState {
/// how many plies left to search in this call
depth: usize,
/// best score (absolute, from current player perspective) guaranteed for current player.
alpha: Option<EvalInt>,
/// best score (absolute, from current player perspective) guaranteed for other player.
beta: Option<EvalInt>,
/// quiescence search flag
quiesce: bool,
}
/// Search the game tree to find the absolute (positive good) move and corresponding eval for the
/// current player.
///
/// This also integrates quiescence search, which looks for a calm (quiescent) position where
/// there are no recaptures.
///
/// # Arguments
///
/// * board: board position to analyze.
/// * depth: how deep to analyze the game tree.
///
/// # Returns
///
/// The best line (in reverse move order), and its corresponding absolute eval for the current player.
fn minmax(board: &mut Board, state: &mut EngineState, mm: MinmaxState) -> (Vec<Move>, SearchEval) {
// these operations are relatively expensive, so only run them occasionally
if state.node_count % (1 << 16) == 0 {
// respect the hard stop if given
match state.rx_engine.try_recv() {
Ok(msg) => match msg {
MsgToEngine::Go(_) => panic!("received go while thinking"),
MsgToEngine::Stop => {
return (Vec::new(), SearchEval::Stopped);
}
MsgToEngine::NewGame => panic!("received newgame while thinking"),
},
Err(e) => match e {
mpsc::TryRecvError::Empty => {}
mpsc::TryRecvError::Disconnected => panic!("thread Main stopped"),
},
}
if let Some(hard) = state.time_lims.hard {
if Instant::now() > hard {
return (Vec::new(), SearchEval::Stopped);
}
}
}
let is_repetition_draw = board.history.count(board.zobrist) >= 2;
let phase_factor = EvalInt::from(board.eval.min_maj_pieces / 5);
// positive here since we're looking from the opposite perspective.
// if white caused a draw, then we'd be black here.
// therefore, white would see a negative value for the draw.
let contempt = state.config.contempt * phase_factor;
// quiescence stand-pat score (only calculated if needed).
// this is where static eval goes.
let mut board_eval: Option<EvalInt> = None;
if mm.quiesce {
board_eval = Some(if is_repetition_draw {
contempt
} else {
board.eval() * EvalInt::from(board.turn.sign())
});
}
if mm.depth == 0 {
if mm.quiesce {
// we hit the limit on quiescence depth
return (Vec::new(), SearchEval::Exact(board_eval.unwrap()));
} else {
// enter quiescence search
return minmax(
board,
state,
MinmaxState {
depth: state.config.qdepth,
alpha: mm.alpha,
beta: mm.beta,
quiesce: true,
},
);
}
}
// default to worst, then gradually improve
let mut alpha = mm.alpha.unwrap_or(EVAL_WORST);
// our best is their worst
let beta = mm.beta.unwrap_or(EVAL_BEST);
let mvs = if mm.quiesce {
board.gen_captures().into_iter().collect::<Vec<_>>()
} else {
board.gen_moves().into_iter().collect::<Vec<_>>()
};
let mut mvs: Vec<_> = mvs
.into_iter()
.map(|mv| (move_priority(board, &mv, state), mv))
.collect();
// get transposition table entry
if state.config.enable_trans_table {
if let Some(entry) = &state.cache[board.zobrist] {
if entry.is_qsearch == mm.quiesce && entry.depth >= mm.depth {
if let SearchEval::Exact(_) | SearchEval::Upper(_) = entry.eval {
// no point looking for a better move
return (vec![entry.best_move], entry.eval);
}
}
mvs.push((EVAL_BEST, entry.best_move));
}
}
// sort moves by decreasing priority
mvs.sort_unstable_by_key(|mv| -mv.0);
let mut abs_best = SearchEval::Exact(EVAL_WORST);
if mm.quiesce {
// stand pat
abs_best = SearchEval::Exact(board_eval.unwrap());
}
let mut best_move: Option<Move> = None;
let mut best_continuation: Vec<Move> = Vec::new();
// determine moves that are allowed in quiescence
if mm.quiesce {
// use static exchange evaluation to prune moves
mvs.retain(|(_priority, mv): &(EvalInt, Move)| -> bool {
let see = board.eval_see(mv.dest, board.turn);
see >= 0
});
}
if mvs.is_empty() {
if mm.quiesce {
// use stand pat
return (Vec::new(), SearchEval::Exact(board_eval.unwrap()));
}
let is_in_check = board.is_check(board.turn);
if is_in_check {
return (Vec::new(), SearchEval::Checkmate(-1));
} else {
// stalemate
return (Vec::new(), SearchEval::Exact(0));
}
}
for (_priority, mv) in mvs {
let anti_mv = mv.make(board);
let (continuation, score) = minmax(
board,
state,
MinmaxState {
depth: mm.depth - 1,
alpha: Some(-beta),
beta: Some(-alpha),
quiesce: mm.quiesce,
},
);
// propagate hard stops
if matches!(score, SearchEval::Stopped) {
return (Vec::new(), SearchEval::Stopped);
}
let abs_score = score.increment();
if abs_score > abs_best {
abs_best = abs_score;
best_move = Some(mv);
best_continuation = continuation;
}
alpha = max(alpha, abs_best.into());
anti_mv.unmake(board);
if alpha >= beta && state.config.alpha_beta_on {
// alpha-beta prune.
//
// Beta represents the best eval that the other player can get in sibling branches
// (different moves in the parent node). Alpha > beta means the eval here is _worse_
// for the other player, so they will never make the move that leads into this branch.
// Therefore, we stop evaluating this branch at all.
if let SearchEval::Upper(eval) | SearchEval::Exact(eval) = abs_best {
abs_best = SearchEval::Lower(eval);
}
break;
}
}
if is_repetition_draw {
abs_best = SearchEval::Exact(contempt);
}
if let Some(best_move) = best_move {
best_continuation.push(best_move);
if state.config.enable_trans_table {
state.cache[board.zobrist] = Some(TranspositionEntry {
best_move,
eval: abs_best,
depth: mm.depth,
is_qsearch: mm.quiesce,
});
}
}
state.node_count += 1;
(best_continuation, abs_best)
}
#[derive(Clone, Copy, Debug)]
pub struct TranspositionEntry {
/// best move found last time
best_move: Move,
/// last time's eval
eval: SearchEval,
/// depth of this entry
depth: usize,
/// is this score within the context of quiescence
is_qsearch: bool,
}
pub type TranspositionTable = ZobristTable<TranspositionEntry>;
/// Iteratively deepen search until it is stopped.
fn iter_deep(board: &mut Board, state: &mut EngineState) -> (Vec<Move>, SearchEval) {
let (mut prev_line, mut prev_eval) = minmax(
board,
state,
MinmaxState {
depth: 1,
alpha: None,
beta: None,
quiesce: false,
},
);
for depth in 2..=state.config.depth {
let (line, eval) = minmax(
board,
state,
MinmaxState {
depth,
alpha: None,
beta: None,
quiesce: false,
},
);
if matches!(eval, SearchEval::Stopped) {
return (prev_line, prev_eval);
} else {
if let Some(soft_lim) = state.time_lims.soft {
if Instant::now() > soft_lim {
return (line, eval);
}
}
(prev_line, prev_eval) = (line, eval);
}
}
(prev_line, prev_eval)
}
/// Deadlines for the engine to think of a move.
#[derive(Default)]
pub struct TimeLimits {
/// The engine must respect this time limit. It will abort if this deadline is passed.
pub hard: Option<Instant>,
pub soft: Option<Instant>,
}
impl TimeLimits {
/// Make time limits based on wtime, btime (but color-independent).
///
/// Also takes in eval metrics, for instance to avoid wasting too much time in the opening.
pub fn from_ourtime_theirtime(ourtime_ms: u64, _theirtime_ms: u64, eval: EvalMetrics) -> Self {
// hard timeout (max)
let mut hard_ms = 100_000;
// soft timeout (default max)
let mut soft_ms = 1_200;
// in some situations we can think longer
if eval.phase <= 13 {
// phase 13 is a single capture of a minor/major piece, so consider that out of the
// opening
soft_ms = if ourtime_ms > 300_000 {
4_500
} else if ourtime_ms > 600_000 {
8_000
} else if ourtime_ms > 1_200_000 {
12_000
} else {
soft_ms
}
}
let factor = if ourtime_ms > 5_000 { 10 } else { 40 };
hard_ms = min(ourtime_ms / factor, hard_ms);
soft_ms = min(ourtime_ms / 50, soft_ms);
let hard_limit = Instant::now() + Duration::from_millis(hard_ms);
let soft_limit = Instant::now() + Duration::from_millis(soft_ms);
TimeLimits {
hard: Some(hard_limit),
soft: Some(soft_limit),
}
}
/// Make time limit based on an exact hard limit.
pub fn from_movetime(movetime_ms: u64) -> Self {
let hard_limit = Instant::now() + Duration::from_millis(movetime_ms);
TimeLimits {
hard: Some(hard_limit),
soft: None,
}
}
}
/// Helper type to avoid retyping the same arguments into every function prototype.
///
/// This should be owned outside the actual thinking part so that the engine can remember state
/// between moves.
pub struct EngineState {
pub config: SearchConfig,
/// Main -> Engine channel receiver
pub rx_engine: mpsc::Receiver<MsgToEngine>,
pub cache: TranspositionTable,
/// Nodes traversed (i.e. number of times minmax called)
pub node_count: usize,
pub time_lims: TimeLimits,
}
impl EngineState {
pub fn new(
config: SearchConfig,
interface: mpsc::Receiver<MsgToEngine>,
cache: TranspositionTable,
time_lims: TimeLimits,
) -> Self {
Self {
config,
rx_engine: interface,
cache,
node_count: 0,
time_lims,
}
}
/// Wipe state between different games.
///
/// Configuration is preserved.
pub fn wipe_state(&mut self) {
self.cache = TranspositionTable::new(self.config.transposition_size);
self.node_count = 0;
}
}
/// Find the best line (in reverse order) and its evaluation.
pub fn best_line(board: &mut Board, engine_state: &mut EngineState) -> (Vec<Move>, SearchEval) {
let (line, eval) = iter_deep(board, engine_state);
(line, eval)
}
/// Find the best move.
pub fn best_move(board: &mut Board, engine_state: &mut EngineState) -> Option<Move> {
let (line, _eval) = best_line(board, engine_state);
line.last().copied()
}
/// Utility for NNUE training set generation to determine if a position is quiet or not.
///
/// Our definition of "quiet" is that there are no checks, and the static and quiescence search
/// evaluations are similar. (See https://arxiv.org/html/2412.17948v1.)
///
/// It is the caller's responsibility to get the search evaluation and pass it to this function.
pub fn is_quiescent_position(board: &Board, eval: SearchEval) -> bool {
// max centipawn value difference to call "similar"
const THRESHOLD: EvalInt = 170;
if board.is_check(board.turn) {
return false;
}
if matches!(eval, SearchEval::Checkmate(_)) {
return false;
}
// white perspective
let abs_eval = EvalInt::from(eval) * EvalInt::from(board.turn.sign());
(board.eval() - EvalInt::from(abs_eval)).abs() <= THRESHOLD.abs()
}