-
Notifications
You must be signed in to change notification settings - Fork 2
/
ddpg_mtcar.py
236 lines (182 loc) · 6.66 KB
/
ddpg_mtcar.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
import time
import argparse
import gym
import numpy as np
from collections import deque
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.nn.init as init
from drawnow import drawnow
import matplotlib.pyplot as plt
last_score_plot = [0]
avg_score_plot = [0]
def draw_fig():
plt.title('reward')
plt.plot(last_score_plot, '-')
plt.plot(avg_score_plot, 'r-')
parser = argparse.ArgumentParser(description='PyTorch DDPG solution of MountainCarContinuous-V0')
parser.add_argument('--gamma', type=float, default=0.99)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--tau', type=float, default=0.001)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--max_episode', type=int, default=200)
parser.add_argument('--max_explore_eps', type=int, default=200)
cfg = parser.parse_args()
class Memory(object):
def __init__(self, memory_size=10000):
self.memory = deque(maxlen=memory_size)
self.memory_size = memory_size
def __len__(self):
return len(self.memory)
def append(self, item):
self.memory.append(item)
def sample_batch(self, batch_size):
idx = np.random.permutation(len(self.memory))[:batch_size]
return [self.memory[i] for i in idx]
# Simple Ornstein-Uhlenbeck Noise generator
# Reference: https://github.com/rllab/rllab/blob/master/rllab/exploration_strategies/ou_strategy.py
def OUNoise():
theta = 0.15
sigma = 0.3
mu = 0
state = 0
while True:
yield state
state += theta * (mu - state) + sigma * np.random.randn()
class Actor(nn.Module):
def __init__(self):
super(Actor, self).__init__()
self.fc_1 = nn.Linear(2, 64)
self.fc_2 = nn.Linear(64, 32)
self.fc_out = nn.Linear(32, 1, bias=False)
init.xavier_normal_(self.fc_1.weight)
init.xavier_normal_(self.fc_2.weight)
init.xavier_normal_(self.fc_out.weight)
def forward(self, x):
out = F.elu(self.fc_1(x))
out = F.elu(self.fc_2(out))
out = F.tanh(self.fc_out(out))
return out
class Critic(nn.Module):
def __init__(self):
super(Critic, self).__init__()
self.fc_state = nn.Linear(2, 32)
self.fc_action = nn.Linear(1, 32)
self.fc = nn.Linear(64, 128)
self.fc_value = nn.Linear(128, 1, bias=False)
init.xavier_normal_(self.fc_state.weight)
init.xavier_normal_(self.fc_action.weight)
init.xavier_normal_(self.fc.weight)
init.xavier_normal_(self.fc_value.weight)
def forward(self, state, action):
out_s = F.elu(self.fc_state(state))
out_a = F.elu(self.fc_action(action))
out = torch.cat([out_s, out_a], dim=1)
out = F.elu(self.fc(out))
out = self.fc_value(out)
return out
def get_action(_actor, state):
if not isinstance(state, torch.Tensor):
state = torch.from_numpy(state).float().cuda()
action = _actor(state)
action = torch.clamp(action, float(env.action_space.low[0]), float(env.action_space.high[0]))
return action
def get_q_value(_critic, state, action):
if not isinstance(state, torch.Tensor):
state = torch.from_numpy(state).float().cuda()
if not isinstance(action, torch.Tensor):
action = torch.from_numpy(action).float().cuda()
q_value = _critic(state, action)
return q_value
def update_actor(state):
action = actor(state)
action = torch.clamp(action, float(env.action_space.low[0]), float(env.action_space.high[0]))
# using chain rule to calculate the gradients of actor
q_value = -torch.mean(critic(state, action))
actor_optimizer.zero_grad()
q_value.backward()
actor_optimizer.step()
return
def update_critic(state, action, target):
q_value = critic(state, action)
loss = F.mse_loss(q_value, target)
critic_optimizer.zero_grad()
loss.backward()
critic_optimizer.step()
return
def soft_update(target, source, tau):
for target_param, param in zip(target.parameters(), source.parameters()):
target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
env = gym.make('MountainCarContinuous-v0')
actor = Actor().cuda()
critic = Critic().cuda()
actor_target = Actor().cuda()
critic_target = Critic().cuda()
actor_target.load_state_dict(actor.state_dict())
critic_target.load_state_dict(critic.state_dict())
actor_optimizer = optim.Adam(actor.parameters(), lr=cfg.lr)
critic_optimizer = optim.Adam(critic.parameters(), lr=cfg.lr)
def main():
# env = wrappers.Monitor(env,'./tmp/',force=True)
state = env.reset()
noise = OUNoise()
iteration_now = 0
iteration = 0
episode = 0
episode_score = 0
episode_steps = 0
memory_warmup = cfg.batch_size * 3
memory = Memory(memory_size=10000)
start_time = time.perf_counter()
while episode < cfg.max_episode:
print('\riter {}, ep {}'.format(iteration_now, episode), end='')
action = get_action(actor, state).item()
# blend determinstic action with random action during exploration
if episode < cfg.max_explore_eps:
p = episode / cfg.max_explore_eps
action = action * p + (1 - p) * next(noise)
next_state, reward, done, _ = env.step([action])
memory.append([state, action, reward, next_state, done])
if iteration >= memory_warmup:
memory_batch = memory.sample_batch(cfg.batch_size)
state_batch, \
action_batch, \
reward_batch, \
next_state_batch, \
done_batch = map(lambda x: torch.tensor(x).float().cuda(), zip(*memory_batch))
action_next = get_action(actor_target, next_state_batch)
# using discounted reward as target q-value to update critic
Q_next = get_q_value(critic_target, next_state_batch, action_next).detach()
Q_target_batch = reward_batch[:, None] + cfg.gamma * (1 - done_batch[:, None]) * Q_next
update_critic(state_batch, action_batch[:, None], Q_target_batch)
# the action corresponds to the state_batch now is nolonger the action stored in buffer,
# so we need to use actor to compute the action first, then use the critic to compute the q-value
update_actor(state_batch)
# soft update
soft_update(actor_target, actor, cfg.tau)
soft_update(critic_target, critic, cfg.tau)
episode_score += reward
episode_steps += 1
iteration_now += 1
iteration += 1
if done:
print(', score {:8f}, steps {}, ({:2f} sec/eps)'.
format(episode_score, episode_steps, time.perf_counter() - start_time))
avg_score_plot.append(avg_score_plot[-1] * 0.99 + episode_score * 0.01)
last_score_plot.append(episode_score)
drawnow(draw_fig)
start_time = time.perf_counter()
episode += 1
episode_score = 0
episode_steps = 0
iteration_now = 0
state = env.reset()
noise = OUNoise()
else:
state = next_state
env.close()
if __name__ == '__main__':
main()
plt.pause(0)