-
Notifications
You must be signed in to change notification settings - Fork 5
/
MADDPG_agent.py
296 lines (253 loc) · 12.5 KB
/
MADDPG_agent.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
296
import random
import copy
from collections import namedtuple, deque
import numpy as np
import torch
import torch.nn.functional as F
import torch.optim as optim
import model
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class MADDPG():
"""Meta agent that contains the two DDPG agents and shared replay buffer."""
def __init__(self, action_size=2, seed=0,
n_agents=2,
buffer_size=10000,
batch_size=256,
gamma=0.99,
update_every=2,
noise_start=1.0,
noise_decay=1.0,
t_stop_noise=30000):
"""
Params
======
action_size (int): dimension of each action
seed (int): Random seed
n_agents (int): number of distinct agents
buffer_size (int): replay buffer size
batch_size (int): minibatch size
gamma (float): discount factor
noise_start (float): initial noise weighting factor
noise_decay (float): noise decay rate
update_every (int): how often to update the network
t_stop_noise (int): max number of timesteps with noise applied in training
"""
self.buffer_size = buffer_size
self.batch_size = batch_size
self.update_every = update_every
self.gamma = gamma
self.n_agents = n_agents
self.noise_weight = noise_start
self.noise_decay = noise_decay
self.t_step = 0
self.noise_on = True
self.t_stop_noise = t_stop_noise
# create two agents, each with their own actor and critic
models = [model.Actor_Critic_Models(n_agents=n_agents) for _ in range(n_agents)]
self.agents = [DDPG(i, models[i]) for i in range(n_agents)]
# create shared replay buffer
self.memory = ReplayBuffer(action_size, self.buffer_size, self.batch_size, seed)
def step(self, all_states, all_actions, all_rewards, all_next_states, all_dones):
all_states = all_states.reshape(1, -1) # reshape 2x24 into 1x48 dim vector
all_next_states = all_next_states.reshape(1, -1) # reshape 2x24 into 1x48 dim vector
self.memory.add(all_states, all_actions, all_rewards, all_next_states, all_dones)
# if t_stop_noise time steps are achieved turn off noise
if self.t_step > self.t_stop_noise:
self.noise_on = False
self.t_step = self.t_step + 1
# Learn every update_every time steps.
if self.t_step % self.update_every == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > self.batch_size:
# sample from the replay buffer for each agent
experiences = [self.memory.sample() for _ in range(self.n_agents)]
self.learn(experiences, self.gamma)
def act(self, all_states, add_noise=True):
# pass each agent's state from the environment and calculate its action
all_actions = []
for agent, state in zip(self.agents, all_states):
action = agent.act(state, noise_weight=self.noise_weight, add_noise=self.noise_on)
self.noise_weight *= self.noise_decay
all_actions.append(action)
return np.array(all_actions).reshape(1, -1) # reshape 2x2 into 1x4 dim vector
def learn(self, experiences, gamma):
# each agent uses its own actor to calculate next_actions
all_next_actions = []
all_actions = []
for i, agent in enumerate(self.agents):
states, _, _, next_states, _ = experiences[i]
agent_id = torch.tensor([i]).to(device)
# extract agent i's state and get action via actor network
state = states.reshape(-1, 2, 24).index_select(1, agent_id).squeeze(1)
action = agent.actor_local(state)
all_actions.append(action)
# extract agent i's next state and get action via target actor network
next_state = next_states.reshape(-1, 2, 24).index_select(1, agent_id).squeeze(1)
next_action = agent.actor_target(next_state)
all_next_actions.append(next_action)
# each agent learns from its experience sample
for i, agent in enumerate(self.agents):
agent.learn(i, experiences[i], gamma, all_next_actions, all_actions)
def save_agents(self):
# save models for local actor and critic of each agent
for i, agent in enumerate(self.agents):
torch.save(agent.actor_local.state_dict(), f"checkpoint_actor_agent_{i}.pth")
torch.save(agent.critic_local.state_dict(), f"checkpoint_critic_agent_{i}.pth")
class DDPG():
"""DDPG agent with own actor and critic."""
def __init__(self, agent_id, model, action_size=2, seed=0,
tau=1e-3,
lr_actor=1e-4,
lr_critic=1e-3,
weight_decay=0.0):
"""
Params
======
model: model object
action_size (int): dimension of each action
seed (int): Random seed
tau (float): for soft update of target parameters
lr_actor (float): learning rate for actor
lr_critic (float): learning rate for critic
weight_decay (float): L2 weight decay
"""
random.seed(seed)
self.id = agent_id
self.action_size = action_size
self.tau = tau
self.lr_actor = lr_actor
self.lr_critic = lr_critic
# Actor Network
self.actor_local = model.actor_local
self.actor_target = model.actor_target
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=lr_actor)
# Critic Network
self.critic_local = model.critic_local
self.critic_target = model.critic_target
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=lr_critic, weight_decay=weight_decay)
# Set weights for local and target actor, respectively, critic the same
self.hard_copy_weights(self.actor_target, self.actor_local)
self.hard_copy_weights(self.critic_target, self.critic_local)
# Noise process
self.noise = OUNoise(action_size, seed)
def hard_copy_weights(self, target, source):
""" copy weights from source to target network (part of initialization)"""
for target_param, param in zip(target.parameters(), source.parameters()):
target_param.data.copy_(param.data)
def act(self, state, noise_weight=1.0, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
# calculate action values
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
self.noise_val = self.noise.sample() * noise_weight
action += self.noise_val
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn(self, agent_id, experiences, gamma, all_next_actions, all_actions):
"""Update policy and value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
all_next_actions (list): each agent's next_action (as calculated by its actor)
all_actions (list): each agent's action (as calculated by its actor)
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# get predicted next-state actions and Q values from target models
self.critic_optimizer.zero_grad()
agent_id = torch.tensor([agent_id]).to(device)
actions_next = torch.cat(all_next_actions, dim=1).to(device)
with torch.no_grad():
q_targets_next = self.critic_target(next_states, actions_next)
# compute Q targets for current states (y_i)
q_expected = self.critic_local(states, actions)
# q_targets = reward of this timestep + discount * Q(st+1,at+1) from target network
q_targets = rewards.index_select(1, agent_id) + (gamma * q_targets_next * (1 - dones.index_select(1, agent_id)))
# compute critic loss
critic_loss = F.mse_loss(q_expected, q_targets.detach())
# minimize loss
critic_loss.backward()
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# compute actor loss
self.actor_optimizer.zero_grad()
# detach actions from other agents
actions_pred = [actions if i == self.id else actions.detach() for i, actions in enumerate(all_actions)]
actions_pred = torch.cat(actions_pred, dim=1).to(device)
actor_loss = -self.critic_local(states, actions_pred).mean()
# minimize loss
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic_local, self.critic_target, self.tau)
self.soft_update(self.actor_local, self.actor_target, self.tau)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)
class OUNoise:
"""Ornstein-Uhlenbeck process."""
def __init__(self, size, seed, mu=0., theta=0.15, sigma=0.2):
"""Initialize parameters and noise process."""
random.seed(seed)
np.random.seed(seed)
self.size = size
self.mu = mu * np.ones(size)
self.theta = theta
self.sigma = sigma
self.reset()
def reset(self):
"""Reset the internal state (= noise) to mean (mu)."""
self.state = copy.copy(self.mu)
def sample(self):
"""Update internal state and return it as a noise sample."""
x = self.state
dx = self.theta * (self.mu - x) + self.sigma * np.random.randn(self.size)
self.state = x + dx
return self.state
class ReplayBuffer():
"""Fixed-size buffer to store experience tuples."""
def __init__(self, action_size, buffer_size, batch_size, seed):
"""Initialize a ReplayBuffer object.
Params
======
action_size (int): dimension of each action
buffer_size (int): maximum size of buffer
batch_size (int): size of each training batch
seed (int): Random seed
"""
random.seed(seed)
np.random.seed(seed)
self.action_size = action_size
self.memory = deque(maxlen=buffer_size)
self.batch_size = batch_size
self.experience = namedtuple("Experience", field_names=["state", "action", "reward", "next_state", "done"])
def add(self, state, action, reward, next_state, done):
"""Add a new experience to memory."""
e = self.experience(state, action, reward, next_state, done)
self.memory.append(e)
def sample(self):
"""Randomly sample a batch of experiences from memory."""
experiences = random.sample(self.memory, k=self.batch_size)
states = torch.from_numpy(np.vstack([e.state for e in experiences if e is not None])).float().to(device)
actions = torch.from_numpy(np.vstack([e.action for e in experiences if e is not None])).float().to(device)
rewards = torch.from_numpy(np.vstack([e.reward for e in experiences if e is not None])).float().to(device)
next_states = torch.from_numpy(np.vstack([e.next_state for e in experiences if e is not None])).float().to(device)
dones = torch.from_numpy(np.vstack([e.done for e in experiences if e is not None]).astype(np.uint8)).float().to(device)
return (states, actions, rewards, next_states, dones)
def __len__(self):
"""Return the current size of internal memory."""
return len(self.memory)