-
Notifications
You must be signed in to change notification settings - Fork 0
/
deep_q.py
295 lines (253 loc) · 12.7 KB
/
deep_q.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
"""
File: deep_q.py
Last update: 03/17/23 by Michelle
Attributions: I followed the PyTorch deep-Q tutorial (https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html),
but typed all code myself and modified it for the Hearts environment
Contains code for the Deep Q-Learning Hearts agent
Author: Michelle Fu
"""
import sys
import math
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
from tqdm import tqdm
import random
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from collections import namedtuple, deque
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
from classes import Card, Trick, Player
from utils import action_from_tsr, state_to_vec, CARD_TO_IND
from simulate_transition import get_starting_state, simulate_transition # comment this out to play against DQN
from baseline_agents import BaselineAgent, GreedyBaseline
from evaluate import evaluate # comment this out to play against DQN
Transition = namedtuple('Transition', ('state', 'action', 'next_state', 'reward', 'legal_mask'))
NUM_PLAYERS = 4
AGENT_INDEX = 0
# stores transitions for experience replay
class ReplayMemory(object):
def __init__(self, capacity):
self.memory = deque([], maxlen=capacity)
# save a transition
def push(self, *args):
self.memory.append(Transition(*args))
# sample a transition
def sample(self, batch_size):
return random.sample(self.memory, batch_size)
def __len__(self):
return len(self.memory)
class DQN(nn.Module):
def __init__(self, n_observations, n_actions) -> None:
super(DQN, self).__init__()
self.layer1 = nn.Linear(n_observations, 256)
self.layer2 = nn.Linear(256, 256)
self.layer3 = nn.Linear(256, n_actions)
# called on one element (to determine next action) or batch
def forward(self, x):
x = F.relu(self.layer1(x))
x = F.relu(self.layer2(x))
return self.layer3(x)
class Trainer():
def __init__(self) -> None:
self.batch_size = 800
self.gamma = 1
self.eps_start = 0.9
self.eps_end = 0.05
self.eps_decay = 1000
self.tau = 0.0001 # update rate of target network
self.lr = 0.00001
# needed for plotting
self.win_percent_b = []
self.lose_percent_b = []
self.orw_p_b = []
self.orl_p_b = []
self.win_percent_g = []
self.lose_percent_g = []
self.orw_p_g = []
self.orl_p_g = []
n_actions = 52 # one for each card
# 52 to encode cards in play, 52 to encode cards in hand, 52 to encode what cards have been previously played
# 4 to encode player order, 4 to encode suit of current hand, 4 to encode which players have won points
n_observations = 168
self.policy_net = DQN(n_observations, n_actions).to(device)
self.target_net = DQN(n_observations, n_actions).to(device)
self.target_net.load_state_dict(self.policy_net.state_dict())
self.optimizer = optim.AdamW(self.policy_net.parameters(), lr=self.lr, amsgrad=True)
self.memory = ReplayMemory(1300)
self.steps_done = 0
def plot_stats(self, type):
plt.figure(1)
x_t = torch.tensor([i * 100 for i in range(len(self.win_percent_b))], dtype=torch.int)
if type == 'baseline':
wins_t = torch.tensor(self.win_percent_b, dtype=torch.float)
losses_t = torch.tensor(self.lose_percent_b, dtype=torch.float)
orw_t = torch.tensor(self.orw_p_b, dtype=torch.float)
orl_t = torch.tensor(self.orl_p_b, dtype=torch.float)
elif type == 'greedy':
wins_t = torch.tensor(self.win_percent_g, dtype=torch.float)
losses_t = torch.tensor(self.lose_percent_g, dtype=torch.float)
orw_t = torch.tensor(self.orw_p_g, dtype=torch.float)
orl_t = torch.tensor(self.orl_p_g, dtype=torch.float)
plt.title('Percentage of wins/losses over time')
plt.xlabel('Episode')
plt.ylabel('Percentage')
plt.plot(x_t.numpy(), wins_t.numpy(), label = "wins")
plt.plot(x_t.numpy(), losses_t.numpy(), label = "losses")
plt.plot(x_t.numpy(), orw_t.numpy(), label = "one round wins")
plt.plot(x_t.numpy(), orl_t.numpy(), label = "one round losses")
plt.legend()
plt.savefig("stats" + type + ".png")
plt.clf()
def get_action(self, state, legal_actions):
eps_threshold = self.eps_end + (self.eps_start - self.eps_end) * \
math.exp(-1 * self.steps_done / self.eps_decay)
self.steps_done += 1
# roll RNG
sample = random.random()
if sample > eps_threshold:
with torch.no_grad():
actions = self.policy_net(state).squeeze()
legal_indices = torch.LongTensor([CARD_TO_IND[card.name] for card in legal_actions]).to(device)
legal_moves_vals = torch.index_select(actions, 0, legal_indices)
action_ind = legal_indices[torch.argmax(legal_moves_vals).item()]
suit, rank, id = action_from_tsr(action_ind)
action = Card(rank, suit, id)
action_tsr = torch.tensor([action_ind], device=device, dtype = torch.long)
return action, action_tsr
else:
action = random.choice(legal_actions)
action_tsr = [CARD_TO_IND[action.name]]
return action, torch.tensor(action_tsr, device=device, dtype = torch.long)
def get_max(self, indices, mask):
action_values = self.target_net(indices)
action_values = action_values.masked_fill(mask, -10000)
return action_values.max(1)[0]
def optimize_model(self):
if len(self.memory) < self.batch_size:
return
transitions = self.memory.sample(self.batch_size)
batch = Transition(*zip(*transitions))
non_final_mask = torch.tensor(tuple(map(lambda s: s is not None, batch.next_state)), device=device, dtype=torch.bool)
non_final_next_states = torch.cat([s for s in batch.next_state if s is not None])
state_batch = torch.cat(batch.state)
action_batch = torch.stack(batch.action)
reward_batch = torch.stack(batch.reward)
mask_batch = torch.stack(batch.legal_mask)
state_action_values = self.policy_net(state_batch).gather(1, action_batch)
next_state_values = torch.zeros(self.batch_size, device=device)
with torch.no_grad():
next_state_values[non_final_mask] = self.get_max(non_final_next_states, mask_batch[non_final_mask])
# compute expected Q values
expected_state_action_values = torch.add((next_state_values * self.gamma).unsqueeze(1),reward_batch)
# compute Huber loss
criterion = nn.SmoothL1Loss()
loss = criterion(state_action_values, expected_state_action_values)
self.optimizer.zero_grad()
loss.backward()
# clip gradients
torch.nn.utils.clip_grad_value_(self.policy_net.parameters(), 100)
self.optimizer.step()
def evaluate_performance(self):
q_agent = deepQAgent(0, self.policy_net)
eval_players_b = [q_agent, BaselineAgent(1), BaselineAgent(2), BaselineAgent(3)]
print("Evaluating against baseline...")
one_round_wins, one_round_losses = evaluate(eval_players_b, end_threshold=0, num_evals=1300)
full_game_wins, full_game_losses = evaluate(eval_players_b, end_threshold=100, num_evals=100)
self.win_percent_b.append(full_game_wins[0])
self.lose_percent_b.append(full_game_losses[0])
self.orw_p_b.append(one_round_wins[0])
self.orl_p_b.append(one_round_losses[0])
eval_players_g = [q_agent, GreedyBaseline(1), GreedyBaseline(2), GreedyBaseline(3)]
print("Evaluating against greedy...")
one_round_wins, one_round_losses = evaluate(eval_players_g, end_threshold=0, num_evals=1300)
full_game_wins, full_game_losses = evaluate(eval_players_g, end_threshold=100, num_evals=100)
self.win_percent_g.append(full_game_wins[0])
self.lose_percent_g.append(full_game_losses[0])
self.orw_p_g.append(one_round_wins[0])
self.orl_p_g.append(one_round_losses[0])
def train(self, num_epochs):
for i in tqdm(range(num_epochs)):
if i % 100 == 0:
self.evaluate_performance()
curr_trick = Trick(NUM_PLAYERS)
tricks = []
hearts_broken = False
state, players = get_starting_state()
state = torch.tensor(state, dtype=torch.float32, device=device).unsqueeze(0)
while True: # play a whole game
this_reward = 0
legal_actions = players[AGENT_INDEX].get_legal_moves(curr_trick, hearts_broken)
action, action_tsr = self.get_action(state, legal_actions)
curr_trick, tricks, players, hearts_broken, next_state, reward = \
simulate_transition(curr_trick, tricks, players, hearts_broken, action) # play the trick to the end and get the next state
if next_state != None:
next_state = torch.tensor(next_state, dtype=torch.float32, device=device).unsqueeze(0)
this_reward += reward
reward_tsr = torch.tensor([reward], dtype=torch.float32, device=device)
mask = torch.zeros(52).to(device)
legal_indices = torch.LongTensor([CARD_TO_IND[card.name] for card in legal_actions]).to(device)
mask = ~(mask.scatter(0, legal_indices, 1.).bool())
self.memory.push(state, action_tsr, next_state, reward_tsr, mask)
state = next_state
self.optimize_model()
if state is None:
break
# soft update target net weights
target_net_state_dict = self.target_net.state_dict()
policy_net_state_dict = self.policy_net.state_dict()
for key in policy_net_state_dict:
target_net_state_dict[key] = policy_net_state_dict[key] * self.tau + target_net_state_dict[key] * (1 - self.tau)
self.target_net.load_state_dict(target_net_state_dict)
self.evaluate_performance()
torch.save(self.policy_net, 'deepq-policy.pt')
class deepQAgent(Player):
def __init__(self, pos: int, policy_net: DQN):
super().__init__(pos)
self.policy_net = policy_net
def take_turn(self, trick: Trick, tricks: 'list[Trick]', players: 'list[Player]', hearts_broken: bool) -> Card:
state = state_to_vec(trick=trick, players=players)
state = torch.tensor(state, dtype=torch.float32, device=device)
with torch.no_grad():
legal_actions = players[AGENT_INDEX].get_legal_moves(trick, hearts_broken)
with torch.no_grad():
actions = self.policy_net(state).squeeze()
legal_indices = torch.LongTensor([CARD_TO_IND[card.name] for card in legal_actions]).to(device)
legal_moves_vals = torch.index_select(actions, 0, legal_indices)
action_ind = legal_indices[torch.argmax(legal_moves_vals).item()]
suit, rank, id = action_from_tsr(action_ind)
for c in self.hand:
if c.suit == suit and c.rank == rank:
self.hand.remove(c)
return c
def main():
num_epochs = 200
if len(sys.argv) == 2:
num_epochs = int(sys.argv[1])
trainer = Trainer()
trainer.train(num_epochs)
policy_net = torch.load('deepq-policy.pt')
policy_net = policy_net.to(device)
q_agent = deepQAgent(0, policy_net)
eval_players = [q_agent, BaselineAgent(1), BaselineAgent(2), BaselineAgent(3)]
print("Evaluating against baseline...")
one_round_wins, one_round_losses = evaluate(eval_players, end_threshold=0, num_evals=10000)
full_game_wins, full_game_losses = evaluate(eval_players, end_threshold=100, num_evals=1000)
print("one round wins:", one_round_wins)
print("one round losses:", one_round_losses)
print("full game wins:", full_game_wins)
print("full game losses", full_game_losses)
eval_players = [q_agent, GreedyBaseline(1), GreedyBaseline(2), GreedyBaseline(3)]
print("Evaluating against greedy...")
one_round_wins, one_round_losses = evaluate(eval_players, end_threshold=0, num_evals=10000)
full_game_wins, full_game_losses = evaluate(eval_players, end_threshold=100, num_evals=1000)
print("one round wins:", one_round_wins)
print("one round losses:", one_round_losses)
print("full game wins:", full_game_wins)
print("full game losses", full_game_losses)
trainer.plot_stats('baseline')
trainer.plot_stats('greedy')
if __name__ == '__main__':
main()