othello/src/complexagent.rs

345 lines
10 KiB
Rust

use crate::{
agent::Agent,
board::{Board, Winner},
piece::Piece,
};
use indicatif::{ProgressBar, ProgressIterator, ProgressStyle};
#[derive(Clone, Debug)]
struct Move {
/// `i` position of move
i: usize,
/// `j` position of move
j: usize,
/// [`Board`] state after move is made
board: Board,
/// Current winner of the match
winner: Winner,
/// Index of this move's parent
parent: Option<usize>,
/// Indices of this Move's Children
children: Vec<usize>,
/// Value of this move
value: i64,
color: Piece,
}
impl Move {
fn compute_self_value(&self, agent_color: Piece, depth: i64) -> i64 {
let mut self_value = self.value;
if self.winner == Winner::Player(!agent_color) {
// if this board results in the opponent winning, MAJORLY negatively weigh this move
// NOTE! this branch isn't completely deleted because if so, the bot wouldn't make a move.
// We shouldn't prune branches because we still need to always react to the opponent's moves
self_value = i64::MIN;
} else if self.winner == Winner::Player(agent_color) {
// results in a win for the agent
self_value = i64::MAX;
}
// TODO! handle ties... what should they be valued as? maybe `i64::MAX / 2`?
self_value / depth
}
}
struct FutureMoves {
arena: Vec<Move>,
current_root: Option<usize>,
current_depth: isize,
max_depth: usize,
/// Color w.r.t
agent_color: Piece,
}
impl FutureMoves {
pub const fn new(agent_color: Piece, max_depth: usize) -> Self {
Self {
arena: Vec::new(),
current_root: None,
current_depth: 0,
max_depth,
agent_color,
}
}
/// Generate children for all children of `nodes`
fn extend_layers(&mut self) {
let mut next_nodes: Vec<usize> = (0..self.arena.len())
.filter(|&idx| self.arena[idx].children.is_empty())
.collect();
for _ in self.current_depth..=(self.max_depth as isize) {
let prog_len = next_nodes.len();
next_nodes = next_nodes
.into_iter()
.progress_with(
ProgressBar::new(prog_len as u64).with_style(
ProgressStyle::with_template(
"Generating children: ({pos}/{len}) {per_sec}",
)
.unwrap(),
),
)
.flat_map(|node_idx| {
self.generate_children(
Some(node_idx),
&self.arena[node_idx].board.clone(),
!self.arena[node_idx].color,
)
})
.collect();
}
self.current_depth = self.max_depth as isize;
}
/// Creates children for a parent (`parent`), returns an iterator it's children's indexes
fn generate_children(
&mut self,
parent: Option<usize>,
board: &Board,
color: Piece,
) -> impl Iterator<Item = usize> {
let start_idx = self.arena.len();
let mut new: Vec<Move> =
// use [`Board::all_positions`] here instead of [`Board::possible_moves`]
// because we use [`Board::what_if`] later and we want to reduce calls to [`Board::propegate_from_dry`]
Board::all_positions()
.flat_map(|(i, j)| board.what_if(i, j, color).map(|x| (i, j, x)))
.map(|(i, j, new_board)| Move {
i,
j,
board: new_board,
winner: new_board.game_winner(color),
parent,
children: Vec::new(),
value: new_board.count(self.agent_color) as i64 - new_board.count(!self.agent_color) as i64,
color: parent.map(|idx| !self.arena[idx].color).unwrap_or(self.agent_color)
}).collect();
// we want to keep only the best move of the agent
if color == self.agent_color {
if new.len() > 1 {
// negative, because we want the max value to be at the first index
new.sort_by_key(|x| -x.compute_self_value(self.agent_color, 1));
new.drain(1..);
}
}
self.arena.extend(new);
let new_indices = start_idx..self.arena.len();
if let Some(parent_idx) = parent {
self.arena[parent_idx].children.extend(new_indices.clone());
}
new_indices
}
/// Given an index from `self.arena`, what depth is it at? 1-indexed (ROOT IS AT INDEX 1)
fn depth_of(&self, node_idx: usize) -> usize {
let mut depth = 0;
let mut current = Some(node_idx);
while let Some(parent_idx) = current {
depth += 1;
current = self.arena[parent_idx].parent;
}
depth
}
fn compute_values(&mut self) {
// PERF! doing this filtering here previously visited moves reduces time
// spent in `depth_of` from 27.79% -> 2.9%
let mut visited = vec![false; self.arena.len()];
for depth in (0..=self.current_depth).rev() {
for idx in 0..self.arena.len() {
if visited[idx] {
continue;
} else {
visited[idx] = true;
}
if self.depth_of(idx) != depth as usize {
continue;
}
let self_value = self.arena[idx]
.compute_self_value(self.agent_color, (self.current_depth - depth + 1) as i64);
let children_value = self.arena[idx]
.children
.iter()
.map(|&child| self.arena[child].value)
.sum::<i64>()
.checked_div(self.arena[idx].children.len() as i64)
.unwrap_or(0);
self.arena[idx].value = self_value + children_value;
}
}
}
pub fn best_move(&self) -> Option<(usize, usize)> {
self.current_root
.and_then(|x| {
self.arena[x]
.children
.iter()
.max_by_key(|&&idx| self.arena[idx].value)
})
.inspect(|&&x| {
assert_eq!(
self.arena[x].color, self.agent_color,
"selected move color should be the same as the color of the agent"
);
})
.map(|&x| (self.arena[x].i, self.arena[x].j))
}
/// Updates `FutureMoves` based on the current state of the board
/// The board is supposed to be after the opposing move
pub fn update(&mut self, board: &Board) {
let curr_board = self
.arena
.iter()
.enumerate()
.find(|(_, m)| {
&m.board == board && (m.parent == self.current_root) && self.current_root.is_some()
})
.map(|(idx, _)| idx);
if let Some(curr_board_idx) = curr_board {
self.current_root = Some(curr_board_idx);
self.update_root_idx(curr_board_idx);
} else {
println!("Generating root of FutureMoves");
self.arena.clear();
self.arena.push(Move {
i: 0,
j: 0,
board: *board,
winner: Winner::None,
parent: None,
children: Vec::new(),
value: 0,
color: !self.agent_color,
});
self.update_root_idx(0);
}
}
pub fn update_root_coord(&mut self, i: usize, j: usize) -> bool {
self.arena
.iter()
.enumerate()
.find_map(|(idx, node)| {
(node.parent == self.current_root && node.i == i && node.j == j).then_some(idx)
})
.inspect(|&root| self.update_root_idx(root))
.is_some()
}
fn update_root_idx(&mut self, idx: usize) {
self.current_root = Some(idx);
self.current_depth -= self.depth_of(idx) as isize - 1;
self.prune_unrelated();
self.extend_layers();
self.compute_values();
}
fn prune_unrelated(&mut self) {
let Some(root) = self.current_root else {
return;
};
// make sure `root` doesn't reference another node
self.arena[root].parent = None;
let mut retain = vec![false; self.arena.len()];
// stack is going to be AT MAXIMUM, the size of the array,
// so lets just pre-allocate it
let mut stack: Vec<usize> = Vec::with_capacity(self.arena.len());
stack.push(root);
// traverse children of the current root
while let Some(idx) = stack.pop() {
retain[idx] = true;
stack.extend(self.arena[idx].children.iter());
}
let mut index_map = vec![None; self.arena.len()];
self.arena = retain
.into_iter()
.enumerate() // old_idx
.zip(self.arena.drain(..))
.flat_map(|((old_idx, keep), node)| keep.then_some((old_idx, node))) // filter out unrelated nodes
.enumerate() // new_idx
.map(|(new_idx, (old_idx, mut node))| {
index_map[old_idx] = Some(new_idx);
node.parent = node.parent.and_then(|p| index_map[p]);
node.children.retain_mut(|c| {
if let Some(new_c) = index_map[*c] {
*c = new_c;
true
} else {
false
}
});
node
})
.collect();
self.current_root = index_map[root];
}
}
pub struct ComplexAgent {
color: Piece,
future_moves: FutureMoves,
}
impl ComplexAgent {
pub const fn new(color: Piece) -> Self {
const MAX_DEPTH: usize = 17;
Self {
color,
future_moves: FutureMoves::new(color, MAX_DEPTH),
}
}
}
impl Agent for ComplexAgent {
fn next_move(&mut self, board: &Board) -> Option<(usize, usize)> {
self.future_moves.update(board);
println!("# of moves stored: {}", self.future_moves.arena.len());
self.future_moves.best_move().inspect(|&(i, j)| {
self.future_moves.update_root_coord(i, j);
})
}
fn name(&self) -> &'static str {
"Complex Agent"
}
fn color(&self) -> Piece {
self.color
}
}