-
Notifications
You must be signed in to change notification settings - Fork 1
/
learner.py
219 lines (198 loc) · 7.78 KB
/
learner.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
import random
from copy import deepcopy
import torch
from numpy import mean
from tqdm import tqdm
# envs
from wizard_env import WizardEnv
from util import argmax, RARingBuffer
from game import GameState
#TODO weight sharing between two networks
#TODO fix card predictor?
class Agent:
def __init__(
self,
env,
card_value_predictor,
trick_value_predictor,
card_decider,
trick_decider,
explore_prob=0.5,
discount=0.9,
device=None,
bufsize=10000,
):
self.device = device or torch.device("cuda:0")
self.card_value_predictor = card_value_predictor.to(self.device)
self.trick_value_predictor = trick_value_predictor.to(self.device)
self.card_decider = card_decider.to(self.device)
self.trick_decider = trick_decider.to(self.device)
self.env = env
self.bufsize = bufsize
self.trick_states = RARingBuffer(
(bufsize, env.observation_dim), dtype=torch.float32, device=self.device
)
self.trick_actions = RARingBuffer(
(bufsize, 1), dtype=torch.long, device=self.device
)
self.card_states = RARingBuffer(
(bufsize, env.observation_dim), dtype=torch.float32, device=self.device
)
self.card_actions = RARingBuffer(
(bufsize, 1), dtype=torch.long, device=self.device
)
# TODO two different state scores, for trick and card?
self.trick_state_scores = RARingBuffer(
(bufsize, 1), dtype=torch.float32, device=self.device
)
self.card_state_scores = RARingBuffer(
(bufsize, 1), dtype=torch.float32, device=self.device
)
self.action_mask = torch.ones(
len(self.env.action_space), dtype=bool, device=self.device
)
self.discount = discount
self.explore_prob = explore_prob
self.card_optimizer = torch.optim.Adam(
self.card_value_predictor.parameters(), lr=1e-3
)
self.trick_optimizer = torch.optim.Adam(
self.trick_value_predictor.parameters(), lr=1e-3
)
def run_episode(self):
state = self.env.reset()
done = False
current_rewards = []
guessing_tricks = []
while not done:
guessing_tricks.append(self.env.game_state == GameState.GuessingTricks)
action = self.get_next_action(state, self.explore_prob)
new_state, reward, done, constraint = self.env.step(action)
self.action_mask = constraint
if guessing_tricks[-1]:
self.trick_states.add(state)
self.trick_actions.add(action)
else:
self.card_states.add(state)
self.card_actions.add(action)
state = new_state
current_rewards.append(reward)
# calculate cumulative rewards
# TODO dont play the entirety of the game, but only one match (n cards -> n tricks -> max of 2+n reward)
for i in range(len(current_rewards)):
G = 0
for j in range(i, len(current_rewards)):
G += current_rewards[j] * self.discount ** (j - i) # TODO maybe MCTS
if guessing_tricks[i]:
self.trick_state_scores.add(G)
else:
self.card_state_scores.add(G)
return sum(current_rewards)
def get_next_action(self, state, eps):
if random.uniform(0, 1) < eps:
return random.choice(
[a for a, i in zip(self.env.action_space, self.action_mask) if i]
)
state = torch.tensor(state, dtype=torch.float32, device=self.device)
state = state.reshape(1, -1)
if self.env.game_state == GameState.GuessingTricks:
values = self.trick_decider(state)
else:
values = self.card_decider(state)
out = argmax(values[0], self.action_mask)
return out
def train(self, epochs=20, update_action_decider=True):
losses = []
for optim, predictor, decider, for_tricks in zip(
(self.card_optimizer, self.trick_optimizer),
(self.card_value_predictor, self.trick_value_predictor),
(self.card_decider, self.trick_decider),
(False, True),
):
for _ in range(epochs):
states, actions, scores = self.sample_memory(
size=self.bufsize//8, for_tricks=for_tricks
)
optim.zero_grad()
state_values = predictor(states)
# maybe FIXME:
# are we losing something if we consider only
# the actions that were actually performed?
# consider keeping the other outputs similar
loss = torch.nn.functional.mse_loss(
state_values.take_along_dim(actions, dim=1), scores
)
loss.backward()
optim.step()
# test the decider
with torch.no_grad():
states, actions, scores = self.sample_memory(
size=self.bufsize//4, for_tricks=for_tricks
)
state_values = predictor(states)
decider_loss = torch.nn.functional.mse_loss(
state_values.take_along_dim(actions, dim=1), scores
)
losses.append(decider_loss.item())
if update_action_decider:
decider.load_state_dict(predictor.state_dict())
return losses
def sample_memory(self, size, for_tricks):
idxs = torch.tensor(random.sample(range(self.bufsize), size))
if for_tricks:
states = self.trick_states.gather(idxs)
actions = self.trick_actions.gather(idxs)
rewards = self.trick_state_scores.gather(idxs)
else:
states = self.card_states.gather(idxs)
actions = self.card_actions.gather(idxs)
rewards = self.card_state_scores.gather(idxs)
return states, actions, rewards
def get_model(n_inputs, n_outputs):
next_power_of_two = 2 ** (n_inputs - 1).bit_length()
return torch.nn.Sequential(
torch.nn.Linear(n_inputs, next_power_of_two),
torch.nn.ReLU(),
torch.nn.Linear(next_power_of_two, next_power_of_two),
torch.nn.ReLU(),
torch.nn.Linear(next_power_of_two, next_power_of_two),
torch.nn.ReLU(),
torch.nn.Linear(next_power_of_two, next_power_of_two),
torch.nn.ReLU(),
torch.nn.Linear(next_power_of_two, n_outputs),
)
if __name__ == "__main__":
EPOCHS = 50000
env = WizardEnv(debug=False, n_rounds = 5)
n_inputs = env.observation_dim
n_outputs = len(env.action_space)
state_value_predictor = get_model(n_inputs, n_outputs)
action_decider = get_model(n_inputs, n_outputs)
action_decider.load_state_dict(state_value_predictor.state_dict())
action_decider.requires_grad_(False)
agent = Agent(
env,
state_value_predictor,
deepcopy(state_value_predictor),
action_decider,
deepcopy(action_decider),
)
agent.explore_prob = 1
for _ in tqdm(range(agent.bufsize)):
agent.run_episode()
bar = tqdm(range(EPOCHS))
rewards = []
horizon = 100
agent.explore_prob *= 1
greedy_decay = (0.001) ** (1 / len(bar))
for i in bar:
cumrewards = agent.run_episode()
trick_loss, card_loss = agent.train(10, i % 5 == 0)
# explore prob should be very small at the end
agent.explore_prob *= greedy_decay
rewards += [cumrewards]
bar.set_description(
f"Trick loss: {trick_loss:.3f}, Card loss: {card_loss:.3f}, current reward: {cumrewards:<2.0f} reward mean: {mean(rewards[-horizon:]):.3f}"
)
env.debug_print = print
from IPython import embed; embed()