diff options
Diffstat (limited to 'src/strategy/minimax.rs')
| -rw-r--r-- | src/strategy/minimax.rs | 104 |
1 files changed, 18 insertions, 86 deletions
diff --git a/src/strategy/minimax.rs b/src/strategy/minimax.rs index 2c52127..5a7bbcd 100644 --- a/src/strategy/minimax.rs +++ b/src/strategy/minimax.rs @@ -7,21 +7,15 @@ use std::cmp; use std::ops::*; use time::{Duration, PreciseTime}; -// TODO: Calibrate these weightings somehow? Some sort of generate and sort based on playing against each other? -// What about: -// - Creating a list (mins and maxes) -// - Keep adding a new guess, run against all, and sort the list by fitness. -// - Repeat until list has many values -// - Somehow prioritize sticking new items in based on what's going well? Or maximally different? Keep dividing all the ranges in half? #[derive(Debug, Clone)] pub struct ScoreConfig { - max_health_weight: f32, - total_health_weight: f32, - points_weight: f32, - victory_weight: f32, - snowball_weight: f32, - bomb_weight: f32, - explore_exploit_weight: f32, + pub max_health_weight: f32, + pub total_health_weight: f32, + pub points_weight: f32, + pub victory_weight: f32, + pub snowball_weight: f32, + pub bomb_weight: f32, + pub explore_exploit_weight: f32, } impl Default for ScoreConfig { @@ -38,7 +32,6 @@ impl Default for ScoreConfig { } } -// TODO: Cache results from last round based on player / opponent move and worm positions pub fn choose_move( state: &GameBoard, config: &ScoreConfig, @@ -86,15 +79,15 @@ pub fn choose_move_with_normalized_perf( let _ = expand_tree(&mut root_node, state.clone(), config); } - eprintln!("Number of simulations: {}", root_node.score_sum.visit_count); - for (command, score_sum) in &root_node.player_score_sums[player_index] { - eprintln!( - "{} = {} ({} visits)", - command, - score_sum.avg().val, - score_sum.visit_count - ); - } + // eprintln!("Number of simulations: {}", root_node.score_sum.visit_count); + // for (command, score_sum) in &root_node.player_score_sums[player_index] { + // eprintln!( + // "{} = {} ({} visits)", + // command, + // score_sum.avg().val, + // score_sum.visit_count + // ); + // } best_player_move(&root_node, player_index) } @@ -178,7 +171,6 @@ fn expand_tree(node: &mut Node, mut state: GameBoard, config: &ScoreConfig) -> S if state.outcome != SimulationOutcome::Continue { score(&state, config) } else if let Some(commands) = node.unexplored.pop() { - // TODO: Explore preemptively doing the rollout? state.simulate(commands); let score = score(&state, config); let unexplored = if state.outcome == SimulationOutcome::Continue { @@ -196,8 +188,6 @@ fn expand_tree(node: &mut Node, mut state: GameBoard, config: &ScoreConfig) -> S node.children.insert(commands, new_node); update(node, commands, score); - // TODO: Prune dominated moves - score } else { let commands = choose_existing(node, config); @@ -215,8 +205,8 @@ fn expand_tree(node: &mut Node, mut state: GameBoard, config: &ScoreConfig) -> S } fn move_combos(state: &GameBoard) -> Vec<[Command; 2]> { - let player_moves = pruned_moves(state, 0); - let opponent_moves = pruned_moves(state, 1); + let player_moves = state.pruned_valid_moves(0); + let opponent_moves = state.pruned_valid_moves(1); debug_assert!(!player_moves.is_empty(), "No player moves"); debug_assert!(!opponent_moves.is_empty(), "No opponent moves"); @@ -255,8 +245,6 @@ fn score(state: &GameBoard, config: &ScoreConfig) -> Score { let snowballs = state.players[0].snowballs() as f32 - state.players[1].snowballs() as f32; let bombs = state.players[0].bombs() as f32 - state.players[1].bombs() as f32; - // TODO: None of these attributes give any signal early on. - // TODO: Try adding new features here. Something about board position? Score { val: max_health * config.max_health_weight + total_health * config.total_health_weight @@ -318,59 +306,3 @@ fn update(node: &mut Node, commands: [Command; 2], score: Score) { .or_insert_with(ScoreSum::new) += score; node.score_sum += score; } - -fn pruned_moves(state: &GameBoard, player_index: usize) -> Vec<Command> { - let sim_with_idle_opponent = |cmd| { - let mut idle_commands = [ - Command::new(Action::DoNothing), - Command::new(Action::DoNothing), - ]; - idle_commands[player_index] = cmd; - let mut state_cpy = state.clone(); - state_cpy.simulate(idle_commands); - state_cpy - }; - - let mut do_nothing_state = state.clone(); - do_nothing_state.simulate([ - Command::new(Action::DoNothing), - Command::new(Action::DoNothing), - ]); - - let opponent_index = GameBoard::opponent(player_index); - let my_starting_health = do_nothing_state.players[player_index].health(); - let opponent_starting_health = do_nothing_state.players[opponent_index].health(); - - state - .valid_moves(player_index) - .into_iter() - .filter(|command| { - // TODO: Some of these filters could be done with better - // performance by running them while generating the list - // of valid moves. - - // NB: These rules should pass for doing nothing, otherwise - // we need some other mechanism for sticking in a do - // nothing option. - - let idle_opponent_state = sim_with_idle_opponent(*command); - let hurt_self = idle_opponent_state.players[player_index].health() < my_starting_health; - let hurt_opponent = - idle_opponent_state.players[opponent_index].health() < opponent_starting_health; - let frozen_opponent = idle_opponent_state.players[opponent_index] - .worms - .iter() - .any(|w| w.rounds_until_unfrozen == FREEZE_DURATION); - - let is_select = command.worm.is_some(); - - let is_attack = command.action.is_attack(); - let is_snowball = command.action.is_snowball(); - - !hurt_self - && (!is_select || hurt_opponent) - && (!is_attack || hurt_opponent) - && (!is_snowball || frozen_opponent) - }) - .collect() -} |