From 41d9a1254477fd2a7ecd10bf2573dc4b62660465 Mon Sep 17 00:00:00 2001 From: Justin Worthe Date: Sun, 7 Jul 2019 14:46:26 +0200 Subject: Strategy split to allow creating a minimax with pruning strategy --- src/strategy/minimax.rs | 236 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 236 insertions(+) create mode 100644 src/strategy/minimax.rs (limited to 'src/strategy/minimax.rs') diff --git a/src/strategy/minimax.rs b/src/strategy/minimax.rs new file mode 100644 index 0000000..deba92e --- /dev/null +++ b/src/strategy/minimax.rs @@ -0,0 +1,236 @@ +use crate::command::{Action, Command}; +use crate::game::{GameBoard, SimulationOutcome}; + +use std::cmp; +use std::collections::HashMap; +use std::ops::*; +use time::{Duration, PreciseTime}; + +pub fn choose_move( + state: &GameBoard, + previous_root: Option, + start_time: PreciseTime, + max_time: Duration, +) -> (Command, Node) { + let mut root_node = match previous_root { + None => Node { + score_sum: ScoreSum::new(), + player_score_sums: [HashMap::new(), HashMap::new()], + unexplored: mcts_move_combo(state), + children: HashMap::new(), + }, + Some(mut node) => node + .children + .drain() + .map(|(_k, n)| n) + .find(|n| false) // TODO: Identify the last opponent move to use this cache + .unwrap_or_else(|| { + eprintln!("Previous round did not appear in the cache"); + Node { + state: state.clone(), + score_sum: ScoreSum::new(), + player_score_sums: [HashMap::new(), HashMap::new()], + unexplored: mcts_move_combo(state), + children: HashMap::new(), + } + }), + }; + + while start_time.to(PreciseTime::now()) < max_time { + let _ = mcts(&mut root_node); + } + + eprintln!("Number of simulations: {}", root_node.score_sum.visit_count); + for (command, score_sum) in &root_node.player_score_sums[0] { + eprintln!( + "{} = {} ({} visits)", + command, + score_sum.avg().val, + score_sum.visit_count + ); + } + + let chosen_command = best_player_move(&root_node); + + root_node + .children + .retain(|[c1, _], _| *c1 == chosen_command); + + (chosen_command, root_node) +} + +pub struct Node { + score_sum: ScoreSum, + player_score_sums: [HashMap; 2], + unexplored: Vec<[Command; 2]>, + children: HashMap<[Command; 2], Node>, +} + +#[derive(Clone, Copy, Debug, PartialEq, PartialOrd)] +struct Score { + val: f32, +} + +impl AddAssign for Score { + fn add_assign(&mut self, other: Self) { + self.val = self.val + other.val; + } +} + +impl Div for Score { + type Output = Self; + fn div(self, other: u32) -> Self { + Score { + val: self.val / other as f32, + } + } +} + +impl cmp::Eq for Score {} +impl cmp::Ord for Score { + fn cmp(&self, other: &Score) -> cmp::Ordering { + self.val + .partial_cmp(&other.val) + .unwrap_or(cmp::Ordering::Equal) + } +} + +struct ScoreSum { + sum: Score, + visit_count: u32, +} + +impl ScoreSum { + fn new() -> ScoreSum { + ScoreSum { + sum: Score { val: 0. }, + visit_count: 0, + } + } + fn with_initial(score: Score) -> ScoreSum { + ScoreSum { + sum: score, + visit_count: 1, + } + } + fn avg(&self) -> Score { + self.sum / self.visit_count + } +} + +impl AddAssign for ScoreSum { + fn add_assign(&mut self, other: Score) { + self.sum += other; + self.visit_count = self.visit_count.saturating_add(1); + } +} + +// TODO: Transform this into more of a minimax with AB pruning approach? +fn mcts(node: &mut Node) -> Score { + if node.state.outcome != SimulationOutcome::Continue { + score(&node.state) + } else if let Some(commands) = node.unexplored.pop() { + let mut new_state = node.state.clone(); + new_state.simulate(commands); + let score = score(&new_state); + let unexplored = if new_state.outcome == SimulationOutcome::Continue { + mcts_move_combo(&new_state) + } else { + Vec::new() + }; + + let new_node = Node { + state: new_state, + score_sum: ScoreSum::with_initial(score), + player_score_sums: [HashMap::new(), HashMap::new()], + unexplored, + children: HashMap::new(), + }; + node.children.insert(commands, new_node); + + update(node, commands, score); + score + } else { + let commands = choose_existing(node); + let score = mcts( + node.children + .get_mut(&commands) + .expect("The existing node hasn't been tried yet"), + ); + update(node, commands, score); + score + } +} + +fn mcts_move_combo(state: &GameBoard) -> Vec<[Command; 2]> { + let player_moves = valid_moves(state, 0); + let opponent_moves = valid_moves(state, 1); + debug_assert!(!player_moves.is_empty(), "No player moves"); + debug_assert!(!opponent_moves.is_empty(), "No opponent moves"); + + let mut result = Vec::with_capacity(player_moves.len() * opponent_moves.len()); + for p in &player_moves { + for o in &opponent_moves { + result.push([*p, *o]); + } + } + + result +} + +fn best_player_move(node: &Node) -> Command { + node.player_score_sums[0] + .iter() + .max_by_key(|(_command, score_sum)| score_sum.avg()) + .map(|(command, _score_sum)| *command) + .unwrap_or_else(|| Command::new(Action::DoNothing)) +} + +fn score(state: &GameBoard) -> Score { + Score { + val: match state.outcome { + SimulationOutcome::PlayerWon(0) => 500., + SimulationOutcome::PlayerWon(1) => -500., + _ => (state.players[0].score() - state.players[1].score()) as f32, + }, + } +} + +fn choose_existing(node: &Node) -> [Command; 2] { + [choose_one_existing(node, 0), choose_one_existing(node, 1)] +} + +fn choose_one_existing(node: &Node, player_index: usize) -> Command { + let ln_n = (node.score_sum.visit_count as f32).ln(); + let c = 100.; + let multiplier = if player_index == 0 { 1. } else { -1. }; + node.player_score_sums[player_index] + .iter() + .max_by_key(|(_command, score_sum)| { + (multiplier * (score_sum.avg().val + c * (ln_n / score_sum.visit_count as f32).sqrt())) + as i32 + }) + .map(|(command, _score_sum)| *command) + .unwrap_or_else(|| Command::new(Action::DoNothing)) +} + +fn update(node: &mut Node, commands: [Command; 2], score: Score) { + *node.player_score_sums[0] + .entry(commands[0]) + .or_insert_with(ScoreSum::new) += score; + *node.player_score_sums[1] + .entry(commands[1]) + .or_insert_with(ScoreSum::new) += score; + node.score_sum += score; +} + +fn valid_moves(state: &GameBoard, player_index: usize) -> Vec { + state + .valid_shoot_commands(player_index) + .iter() + .chain(state.valid_move_commands(player_index).iter()) + .chain(state.valid_bomb_commands(player_index).iter()) + .chain([Command::new(Action::DoNothing)].iter()) + .cloned() + .collect() +} -- cgit v1.2.3