Added basic heuristic
[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     let needs_energy = player.energy_generated() <= ENERGY_PRODUCTION_CUTOFF ||
212         player.energy <= ENERGY_STORAGE_CUTOFF;
213     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
214         let i = e_base + p as usize;
215         let point = Point::new_index(p);
216         let weight = if !needs_energy || player.energy < ENERGY_PRICE || player.occupied & point.to_either_bitfield() != 0 {
217             0
218         } else if point.x() < 2 {
219             2
220         } else {
221             1
222         };
223
224         cumulative_distribution += weight;
225         cdf[i] = cumulative_distribution;
226     }
227
228     // Defence
229     let d_base = e_base + NUMBER_OF_MAP_POSITIONS;
230     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
231         let i = d_base + p as usize;
232         let point = Point::new_index(p);
233         let weight = if player.energy < DEFENCE_PRICE || player.occupied & point.to_either_bitfield() != 0 {
234             0
235         } else {
236             //TODO: Favour the front and rows with something to defend
237             opponent.count_attack_towers_in_row(point.y())
238         };
239
240         cumulative_distribution += weight;
241         cdf[i] = cumulative_distribution;
242     }
243
244     // Attack
245     let a_base = d_base + NUMBER_OF_MAP_POSITIONS;
246     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
247         let i = a_base + p as usize;
248         let point = Point::new_index(p);
249         let weight = if  player.energy < MISSILE_PRICE || player.occupied & point.to_either_bitfield() != 0 {
250             0
251         } else {
252             8 + opponent.count_energy_towers_in_row(point.y()) - opponent.count_attack_towers_in_row(point.y())
253         };
254
255         cumulative_distribution += weight;
256         cdf[i] = cumulative_distribution;
257     }
258
259     // Tesla
260     let t_base = a_base + NUMBER_OF_MAP_POSITIONS;
261     let cant_tesla = player.has_max_teslas() || player.energy < TESLA_PRICE;
262     for p in 0..NUMBER_OF_MAP_POSITIONS as u8 {
263         let i = t_base + p as usize;
264         let point = Point::new_index(p);
265         let weight = if cant_tesla || (player.occupied & point.to_either_bitfield() != 0) || point.y() < 7 {
266             0
267         } else {
268             2
269         };
270
271         cumulative_distribution += weight;
272         cdf[i] = cumulative_distribution;
273     }
274
275     assert_eq!(MOVES.len(), t_base + NUMBER_OF_MAP_POSITIONS);
276
277     if cumulative_distribution == 0 {
278         return Command::Nothing;
279     }
280
281     let choice = rng.gen_range(0, cumulative_distribution);
282
283     // find maximum index where choice <= cdf[index]
284     // TODO can this be a more efficient lookup?
285     let index = cdf.iter().position(|&c| c > choice).expect("Random number has exceeded cumulative distribution");
286
287     MOVES[index].clone()
288 }
289
290 #[cfg(not(feature = "heuristic-random"))]
291 fn random_move<R: Rng>(player: &Player, _opponent: &Player, rng: &mut R) -> Command {
292     let free_positions_count = player.unoccupied_cell_count();
293
294     let open_building_spot = free_positions_count > 0;
295
296     let all_buildings = sensible_buildings(player, open_building_spot);
297
298     let iron_curtain_count = if player.can_build_iron_curtain() && player.energy >= IRON_CURTAIN_PRICE { 1 } else { 0 };
299     let nothing_count = 1;
300
301     let building_choice_index = rng.gen_range(0, all_buildings.len() + nothing_count + iron_curtain_count);
302
303     if building_choice_index < all_buildings.len() {
304         let position_choice = rng.gen_range(0, free_positions_count);
305         Command::Build(
306             player.location_of_unoccupied_cell(position_choice),
307             all_buildings[building_choice_index]
308         )
309     }
310     else if building_choice_index == all_buildings.len() {
311         Command::Nothing        
312     } else {
313         Command::IronCurtain
314     }
315 }
316
317 #[derive(Debug)]
318 struct CommandScore {
319     command: Command,
320     starts_with_nothing: bool,
321     victories: u32,
322     defeats: u32,
323     draws: u32,
324     stalemates: u32,
325     attempts: u32,
326     next_seed: [u8; 16]
327 }
328
329 impl CommandScore {
330     fn new(command: Command, starts_with_nothing: bool) -> CommandScore {
331         CommandScore {
332             command, starts_with_nothing,
333             victories: 0,
334             defeats: 0,
335             draws: 0,
336             stalemates: 0,
337             attempts: 0,
338             next_seed: INIT_SEED
339         }
340     }
341
342     fn add_victory(&mut self, next_seed: [u8; 16]) {
343         self.victories += 1;
344         self.attempts += 1;
345         self.next_seed = next_seed;
346     }
347
348     fn add_defeat(&mut self, next_seed: [u8; 16]) {
349         self.defeats += 1;
350         self.attempts += 1;
351         self.next_seed = next_seed;
352     }
353
354     fn add_draw(&mut self, next_seed: [u8; 16]) {
355         self.draws += 1;
356         self.attempts += 1;
357         self.next_seed = next_seed;
358     }
359
360     fn add_stalemate(&mut self, next_seed: [u8; 16]) {
361         self.stalemates += 1;
362         self.attempts += 1;
363         self.next_seed = next_seed;
364     }
365
366     fn win_ratio(&self) -> i32 {
367         (self.victories as i32 - self.defeats as i32) * 10000 / (self.attempts as i32)
368     }
369
370     fn init_command_scores(state: &BitwiseGameState) -> Vec<CommandScore> {
371         let unoccupied_cells_count = state.player.unoccupied_cell_count();
372         let unoccupied_cells = (0..unoccupied_cells_count)
373             .map(|i| state.player.location_of_unoccupied_cell(i));
374         let energy_generated = state.player.energy_generated();
375
376         let mut all_buildings: ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> = ArrayVec::new();
377         if DEFENCE_PRICE <= state.player.energy {
378             all_buildings.push(BuildingType::Defence);
379         }
380         if MISSILE_PRICE <= state.player.energy {
381             all_buildings.push(BuildingType::Attack);
382         }
383         if ENERGY_PRICE <= state.player.energy {
384             all_buildings.push(BuildingType::Energy);
385         }
386         if !state.player.has_max_teslas() && (TESLA_PRICE.saturating_sub(state.player.energy) / energy_generated < 4) {
387             all_buildings.push(BuildingType::Tesla);
388         }
389         
390         let building_command_count = unoccupied_cells.len()*all_buildings.len();
391         
392         let mut commands = Vec::with_capacity(building_command_count + 1);
393         if state.player.can_build_iron_curtain() && IRON_CURTAIN_PRICE.saturating_sub(state.player.energy) / energy_generated < 4 {
394             commands.push(CommandScore::new(Command::IronCurtain, state.player.energy < IRON_CURTAIN_PRICE));
395         }
396
397         for position in unoccupied_cells {
398             for &building in &all_buildings {
399                 commands.push(CommandScore::new(Command::Build(position, building), building.cant_build_yet(state.player.energy)));
400             }
401         }
402
403         commands
404     }
405 }
406
407 impl fmt::Display for CommandScore {
408     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
409         write!(f, "{},{}", self.command, self.win_ratio())
410     }
411 }
412
413 #[cfg(not(feature = "energy-cutoff"))]
414 fn sensible_buildings(player: &Player, open_building_spot: bool) -> ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> {
415     let mut result = ArrayVec::new();
416     if !open_building_spot {
417         return result;
418     }
419
420     if DEFENCE_PRICE <= player.energy {
421         result.push(BuildingType::Defence);
422     }
423     if MISSILE_PRICE <= player.energy {
424         result.push(BuildingType::Attack);
425     }
426     if ENERGY_PRICE <= player.energy {
427         result.push(BuildingType::Energy);
428     }
429     if TESLA_PRICE <= player.energy && !player.has_max_teslas() {
430         result.push(BuildingType::Tesla);
431     }
432
433     result
434 }
435
436 #[cfg(feature = "energy-cutoff")]
437 fn sensible_buildings(player: &Player, open_building_spot: bool) -> ArrayVec<[BuildingType; NUMBER_OF_BUILDING_TYPES]> {
438     let mut result = ArrayVec::new();
439     if !open_building_spot {
440         return result;
441     }
442
443     let needs_energy = player.energy_generated() <= ENERGY_PRODUCTION_CUTOFF ||
444         player.energy <= ENERGY_STORAGE_CUTOFF;
445
446     if DEFENCE_PRICE <= player.energy {
447         result.push(BuildingType::Defence);
448     }
449     if MISSILE_PRICE <= player.energy {
450         result.push(BuildingType::Attack);
451     }
452     if ENERGY_PRICE <= player.energy && needs_energy {
453         result.push(BuildingType::Energy);
454     }
455     if TESLA_PRICE <= player.energy && !player.has_max_teslas() {
456         result.push(BuildingType::Tesla);
457     }
458     
459     result
460 }
461