Improved perf by breaking strong ordering dependency in indexes
[entelect-challenge-tower-defence.git] / src / strategy / monte_carlo.rs
1 use engine::command::*;
2 use engine::status::GameStatus;
3 use engine::bitwise_engine::{Player, BitwiseGameState};
4 use engine::constants::*;
5 use engine::geometry::*;
6
7 use std::fmt;
8
9 use rand::{Rng, XorShiftRng, SeedableRng};
10
11 use arrayvec::ArrayVec;
12
13 const MAX_MOVES: u16 = 400;
14 const INIT_SEED: [u8;16] = [0x7b, 0x6a, 0xe1, 0xf4, 0x41, 0x3c, 0xe9, 0x0f, 0x67, 0x81, 0x67, 0x99, 0x77, 0x0a, 0x6b, 0xda];
15
16 use time::{Duration, PreciseTime};
17
18 #[cfg(not(feature = "single-threaded"))]
19 use rayon::prelude::*;
20
21 #[cfg(feature = "energy-cutoff")] pub const ENERGY_PRODUCTION_CUTOFF: u16 = 50;
22 #[cfg(feature = "energy-cutoff")] pub const ENERGY_STORAGE_CUTOFF: u16 = 120;
23
24 pub fn choose_move(state: &BitwiseGameState, start_time: PreciseTime, max_time: Duration) -> Command {
25     let mut command_scores = CommandScore::init_command_scores(state);
26
27     let command = {
28         let best_command_score = simulate_options_to_timeout(&mut command_scores, state, start_time, max_time);
29         match best_command_score {
30             Some(best) if !best.starts_with_nothing => best.command,
31             _ => Command::Nothing
32         }
33     };
34
35     #[cfg(feature = "benchmarking")]
36     {
37         let total_iterations: u32 = command_scores.iter().map(|c| c.attempts).sum();
38         println!("Iterations: {}", total_iterations);
39     }
40     #[cfg(feature = "debug-decisions")]
41     {
42         debug_print_choices("ENERGY", &command_scores, |score| match score.command {
43             Command::Build(p, BuildingType::Energy) => Some((p, score.win_ratio())),
44             _ => None
45         });
46         debug_print_choices("ATTACK", &command_scores, |score| match score.command {
47             Command::Build(p, BuildingType::Attack) => Some((p, score.win_ratio())),
48             _ => None
49         });
50         debug_print_choices("DEFENCE", &command_scores, |score| match score.command {
51             Command::Build(p, BuildingType::Defence) => Some((p, score.win_ratio())),
52             _ => None
53         });
54         debug_print_choices("TESLA", &command_scores, |score| match score.command {
55             Command::Build(p, BuildingType::Tesla) => Some((p, score.win_ratio())),
56             _ => None
57         });
58         
59         println!("NOTHING");
60         println!("{}", command_scores.iter().find(|c| c.command == Command::Nothing).map(|s| s.win_ratio()).unwrap_or(0));
61         println!("");
62
63         println!("IRON CURTAIN");
64         println!("{}", command_scores.iter().find(|c| c.command == Command::IronCurtain).map(|s| s.win_ratio()).unwrap_or(0));
65         println!("");
66     }
67
68     command
69 }
70
71 #[cfg(feature = "debug-decisions")]
72 fn debug_print_choices<F: FnMut(&CommandScore) -> Option<(Point, i32)>>(label: &str, command_scores: &[CommandScore], extractor: F) {
73     println!("#+NAME: {}", label);
74     println!("#+PLOT: type:3d with:pm3d");
75     let relevant_moves: Vec<(Point, i32)>  = command_scores.iter()
76         .filter_map(extractor)
77         .collect();
78     for y in 0..MAP_HEIGHT {
79         for x in 0..SINGLE_MAP_WIDTH {
80             let point = Point::new(x, y);
81             let score = relevant_moves.iter().find(|(p, _)| *p == point);
82             print!(" | {}", score.map(|(_,s)| s).cloned().unwrap_or(0));
83         }
84         println!(" |");
85     }
86     println!("");
87 }
88
89 #[cfg(not(feature = "discard-poor-performers"))]
90 fn simulate_options_to_timeout(command_scores: &'a mut Vec<CommandScore>, settings: &GameSettings, state: &BitwiseGameState, start_time: PreciseTime, max_time: Duration) -> Option<&'a CommandScore> {
91     loop {
92         simulate_all_options_once(command_scores, settings, state);
93         if start_time.to(PreciseTime::now()) > max_time {
94             break;
95         }
96     }
97     command_scores.iter().max_by_key(|&c| c.win_ratio())
98 }
99
100 #[cfg(feature = "discard-poor-performers")]
101 fn simulate_options_to_timeout<'a>(command_scores: &'a mut Vec<CommandScore>, state: &BitwiseGameState, start_time: PreciseTime, max_time: Duration) -> Option<&'a CommandScore> {
102     use std::cmp;
103     let min_options = cmp::min(command_scores.len(), 5);
104     
105     let maxes = [max_time / 4, max_time / 2, max_time * 3 / 4, max_time];
106     for (i, &max) in maxes.iter().enumerate() {
107         let new_length = cmp::max(min_options, command_scores.len() / (2usize.pow(i as u32)));
108         let active_scores = &mut command_scores[0..new_length];
109         loop {
110             simulate_all_options_once(active_scores, state);
111             if start_time.to(PreciseTime::now()) > max {
112                 break;
113             }
114         }
115         active_scores.sort_unstable_by_key(|c| -c.win_ratio());
116     }
117     command_scores.first()
118 }
119
120 #[cfg(feature = "single-threaded")]
121 fn simulate_all_options_once(command_scores: &mut[CommandScore], state: &BitwiseGameState) {
122     command_scores.iter_mut()
123         .for_each(|score| {
124             let mut rng = XorShiftRng::from_seed(score.next_seed);
125             simulate_to_endstate(score, state, &mut rng);
126         });
127 }
128
129 #[cfg(not(feature = "single-threaded"))]
130 fn simulate_all_options_once(command_scores: &mut[CommandScore], state: &BitwiseGameState) {
131     command_scores.par_iter_mut()
132         .for_each(|score| {
133             let mut rng = XorShiftRng::from_seed(score.next_seed);
134             simulate_to_endstate(score, state, &mut rng);
135         });
136 }
137
138 fn simulate_to_endstate<R: Rng>(command_score: &mut CommandScore, state: &BitwiseGameState, rng: &mut R) {
139     let mut state_mut = state.clone();
140     
141     let mut status = GameStatus::Continue; //state_mut.simulate(command_score.command, opponent_first);
142     let mut first_move_made = false;
143     
144     for _ in 0..MAX_MOVES {
145         if status != GameStatus::Continue {
146             break;
147         }
148
149         let player_command = if first_move_made {
150             random_move(&state_mut.player, &state_mut.opponent, rng)
151         } else {
152             let do_nothing = command_score.command.cant_build_yet(state_mut.player.energy);
153             first_move_made = !do_nothing;
154             if do_nothing { Command::Nothing } else { command_score.command }
155         };
156         let opponent_command = random_move(&state_mut.opponent, &state_mut.player, rng);
157         status = state_mut.simulate(player_command, opponent_command);
158     }
159
160     let mut next_seed: [u8;16] = [0; 16];
161     rng.fill_bytes(&mut next_seed);
162     match status {
163         GameStatus::PlayerWon => command_score.add_victory(next_seed),
164         GameStatus::OpponentWon => command_score.add_defeat(next_seed),
165         GameStatus::Continue => command_score.add_stalemate(next_seed),
166         GameStatus::Draw => command_score.add_draw(next_seed)
167     }
168 }
169
170 #[cfg(feature = "heuristic-random")]
171 fn random_move<R: Rng>(player: &Player, _opponent: &Player, rng: &mut R) -> Command {
172     lazy_static! {
173         static ref MOVES: [Command; NUMBER_OF_POSSIBLE_MOVES] = {
174             let mut m = [Command::Nothing; NUMBER_OF_POSSIBLE_MOVES];
175             m[1] = Command::IronCurtain;
176             let mut i = 2;
177             for b in [BuildingType::Energy, BuildingType::Defence, BuildingType::Attack, BuildingType::Tesla].iter() {
178                 for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
179                 let point = Point::new_index(p);
180                     m[i] = Command::Build(point, *b);
181                     i += 1;
182                 }
183             }
184             m
185         };
186     }
187
188     let mut cdf = [0; NUMBER_OF_POSSIBLE_MOVES];
189     let mut cumulative_distribution: u32 = 0;
190
191     // Nothing
192     {
193         let weight = 0;
194         cumulative_distribution += weight;
195         cdf[0] = cumulative_distribution;
196     }
197     
198     // Iron Curtain
199     {
200         let weight = if player.can_build_iron_curtain() && player.energy >= IRON_CURTAIN_PRICE {
201             20
202         } else {
203             0
204         };
205         cumulative_distribution += weight;
206         cdf[1] = cumulative_distribution;
207     }
208
209     // Energy
210     let e_base = 2;
211     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
212         let i = e_base + p as usize;
213         let point = Point::new_index(p);
214         let weight = if player.energy < ENERGY_PRICE || player.occupied & point.to_either_bitfield() != 0 {
215             0
216         } else {
217             //TODO: This needs to be more complex
218             1
219         };
220
221         cumulative_distribution += weight;
222         cdf[i] = cumulative_distribution;
223     }
224
225     // Defence
226     let d_base = e_base + NUMBER_OF_MAP_POSITIONS;
227     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
228         let i = d_base + p as usize;
229         let point = Point::new_index(p);
230         let weight = if player.energy < DEFENCE_PRICE || player.occupied & point.to_either_bitfield() != 0 {
231             0
232         } else {
233             //TODO: This needs to be more complex
234             1
235         };
236
237         cumulative_distribution += weight;
238         cdf[i] = cumulative_distribution;
239     }
240
241     // Attack
242     let a_base = d_base + NUMBER_OF_MAP_POSITIONS;
243     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
244         let i = a_base + p as usize;
245         let point = Point::new_index(p);
246         let weight = if  player.energy < MISSILE_PRICE || player.occupied & point.to_either_bitfield() != 0 {
247             0
248         } else {
249             //TODO: This needs to be more complex
250             1
251         };
252
253         cumulative_distribution += weight;
254         cdf[i] = cumulative_distribution;
255     }
256
257     // Tesla
258     let t_base = a_base + NUMBER_OF_MAP_POSITIONS;
259     let cant_tesla = player.has_max_teslas() || player.energy < TESLA_PRICE;
260     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
261         let i = t_base + p as usize;
262         let point = Point::new_index(p);
263         let weight = if cant_tesla || player.occupied & point.to_either_bitfield() != 0 {
264             0
265         } else {
266             //TODO: This needs to be more complex
267             1
268         };
269
270         cumulative_distribution += weight;
271         cdf[i] = cumulative_distribution;
272     }
273
274     assert_eq!(MOVES.len(), t_base + NUMBER_OF_MAP_POSITIONS);
275
276     if cumulative_distribution == 0 {
277         return Command::Nothing;
278     }
279
280     let choice = rng.gen_range(0, cumulative_distribution);
281
282     // find maximum index where choice <= cdf[index]
283     let index = cdf.iter().position(|&c| c > choice).expect("Random number has exceeded cumulative distribution");
284     
285     /*
286     let mut min_index = 0;
287     let mut max_index = cdf.len();
288
289     while max_index - min_index > 1 {
290         let mid = (min_index + max_index) / 2;
291         if cdf[mid] > choice {
292             max_index = mid+1;
293         } else {
294             min_index = mid;
295         }
296     }
297     let index = min_index;
298      */
299     
300     MOVES[index].clone()
301 }
302
303 #[cfg(not(feature = "heuristic-random"))]
304 fn random_move<R: Rng>(player: &Player, _opponent: &Player, rng: &mut R) -> Command {
305     let free_positions_count = player.unoccupied_cell_count();
306
307     let open_building_spot = free_positions_count > 0;
308
309     let all_buildings = sensible_buildings(player, open_building_spot);
310
311     let iron_curtain_count = if player.can_build_iron_curtain() && player.energy >= IRON_CURTAIN_PRICE { 1 } else { 0 };
312     let nothing_count = 1;
313
314     let building_choice_index = rng.gen_range(0, all_buildings.len() + nothing_count + iron_curtain_count);
315
316     if building_choice_index < all_buildings.len() {
317         let position_choice = rng.gen_range(0, free_positions_count);
318         Command::Build(
319             player.location_of_unoccupied_cell(position_choice),
320             all_buildings[building_choice_index]
321         )
322     }
323     else if building_choice_index == all_buildings.len() {
324         Command::Nothing        
325     } else {
326         Command::IronCurtain
327     }
328 }
329
330 #[derive(Debug)]
331 struct CommandScore {
332     command: Command,
333     starts_with_nothing: bool,
334     victories: u32,
335     defeats: u32,
336     draws: u32,
337     stalemates: u32,
338     attempts: u32,
339     next_seed: [u8; 16]
340 }
341
342 impl CommandScore {
343     fn new(command: Command, starts_with_nothing: bool) -> CommandScore {
344         CommandScore {
345             command, starts_with_nothing,
346             victories: 0,
347             defeats: 0,
348             draws: 0,
349             stalemates: 0,
350             attempts: 0,
351             next_seed: INIT_SEED
352         }
353     }
354
355     fn add_victory(&mut self, next_seed: [u8; 16]) {
356         self.victories += 1;
357         self.attempts += 1;
358         self.next_seed = next_seed;
359     }
360
361     fn add_defeat(&mut self, next_seed: [u8; 16]) {
362         self.defeats += 1;
363         self.attempts += 1;
364         self.next_seed = next_seed;
365     }
366
367     fn add_draw(&mut self, next_seed: [u8; 16]) {
368         self.draws += 1;
369         self.attempts += 1;
370         self.next_seed = next_seed;
371     }
372
373     fn add_stalemate(&mut self, next_seed: [u8; 16]) {
374         self.stalemates += 1;
375         self.attempts += 1;
376         self.next_seed = next_seed;
377     }
378
379     fn win_ratio(&self) -> i32 {
380         (self.victories as i32 - self.defeats as i32) * 10000 / (self.attempts as i32)
381     }
382
383     fn init_command_scores(state: &BitwiseGameState) -> Vec<CommandScore> {
384         let unoccupied_cells_count = state.player.unoccupied_cell_count();
385         let unoccupied_cells = (0..unoccupied_cells_count)
386             .map(|i| state.player.location_of_unoccupied_cell(i));
387         let energy_generated = state.player.energy_generated();
388
389         let mut all_buildings: ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> = ArrayVec::new();
390         if DEFENCE_PRICE <= state.player.energy {
391             all_buildings.push(BuildingType::Defence);
392         }
393         if MISSILE_PRICE <= state.player.energy {
394             all_buildings.push(BuildingType::Attack);
395         }
396         if ENERGY_PRICE <= state.player.energy {
397             all_buildings.push(BuildingType::Energy);
398         }
399         if !state.player.has_max_teslas() && (TESLA_PRICE.saturating_sub(state.player.energy) / energy_generated < 4) {
400             all_buildings.push(BuildingType::Tesla);
401         }
402         
403         let building_command_count = unoccupied_cells.len()*all_buildings.len();
404         
405         let mut commands = Vec::with_capacity(building_command_count + 1);
406         if state.player.can_build_iron_curtain() && IRON_CURTAIN_PRICE.saturating_sub(state.player.energy) / energy_generated < 4 {
407             commands.push(CommandScore::new(Command::IronCurtain, state.player.energy < IRON_CURTAIN_PRICE));
408         }
409
410         for position in unoccupied_cells {
411             for &building in &all_buildings {
412                 commands.push(CommandScore::new(Command::Build(position, building), building.cant_build_yet(state.player.energy)));
413             }
414         }
415
416         commands
417     }
418 }
419
420 impl fmt::Display for CommandScore {
421     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
422         write!(f, "{},{}", self.command, self.win_ratio())
423     }
424 }
425
426 #[cfg(not(feature = "energy-cutoff"))]
427 fn sensible_buildings(player: &Player, open_building_spot: bool) -> ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> {
428     let mut result = ArrayVec::new();
429     if !open_building_spot {
430         return result;
431     }
432
433     if DEFENCE_PRICE <= player.energy {
434         result.push(BuildingType::Defence);
435     }
436     if MISSILE_PRICE <= player.energy {
437         result.push(BuildingType::Attack);
438     }
439     if ENERGY_PRICE <= player.energy {
440         result.push(BuildingType::Energy);
441     }
442     if TESLA_PRICE <= player.energy && !player.has_max_teslas() {
443         result.push(BuildingType::Tesla);
444     }
445
446     result
447 }
448
449 #[cfg(feature = "energy-cutoff")]
450 fn sensible_buildings(player: &Player, open_building_spot: bool) -> ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> {
451     let mut result = ArrayVec::new();
452     if !open_building_spot {
453         return result;
454     }
455
456     let needs_energy = player.energy_generated() <= ENERGY_PRODUCTION_CUTOFF ||
457         player.energy <= ENERGY_STORAGE_CUTOFF;
458
459     if DEFENCE_PRICE <= player.energy {
460         result.push(BuildingType::Defence);
461     }
462     if MISSILE_PRICE <= player.energy {
463         result.push(BuildingType::Attack);
464     }
465     if ENERGY_PRICE <= player.energy && needs_energy {
466         result.push(BuildingType::Energy);
467     }
468     if TESLA_PRICE <= player.energy && !player.has_max_teslas() {
469         result.push(BuildingType::Tesla);
470     }
471     
472     result
473 }
474