-
Notifications
You must be signed in to change notification settings - Fork 0
/
maddpg.py
202 lines (137 loc) · 9 KB
/
maddpg.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
# main code that contains the neural network setup
# policy + critic updates
# see ddpg.py for other details in the network
from ddpg import DDPGAgent
import torch
from utilities import soft_update, transpose_to_tensor, transpose_list , transpose_to_tensorAsitis , giveCurrentAgentsAction
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#device = 'cpu'
import numpy as np
import torch.nn.functional as F
class MADDPG:
def __init__(self, state_size , action_size , discount_factor=0.95, tau=0.05 , lr_actor = 2e-4 , lr_critic = 2e-3 , num_agents =2):
super(MADDPG, self).__init__()
hidden_in_dim = 512
hidden_out_dim = 256
# critic input = obs_full + actions = 48+2+2=52
# have to change the agent neurons for sure
# the no of agents is two because there are only two players
self.maddpg_agent = [DDPGAgent(state_size , action_size,hidden_in_dim, hidden_out_dim, num_agents =num_agents , lr_actor = lr_actor, lr_critic = lr_critic),
DDPGAgent(state_size , action_size, hidden_in_dim, hidden_out_dim, num_agents =num_agents, lr_actor = lr_actor, lr_critic = lr_critic)]
self.num_agents = num_agents
self.action_vector = 2
self.discount_factor = discount_factor
self.tau = tau
self.iter = 0
def get_actors(self):
"""get target_actors of all """
"""get actors of all the agents in the MADDPG object"""
actors = [ddpg_agent.actor for ddpg_agent in self.maddpg_agent]
return actors
def get_target_actors(self):
"""the agents in the MADDPG object"""
target_actors = [ddpg_agent.target_actor for ddpg_agent in self.maddpg_agent]
return target_actors
def act(self, obs_all_agents, noise=0.0 , batch = True):
"""get actions from all agents in the MADDPG object"""
#print(obs_all_agents)
#shape_vec = [np.shape(obs) for obs in obs_all_agents]
#print("shape",shape_vec)
actions = [agent.act(obs, noise , batch = batch) for agent, obs in zip(self.maddpg_agent, obs_all_agents)]
return actions
def target_act(self,agent_no, obs_all_agents, noise=0.0 , batch = True):
"""get target network actions from all the agents in the MADDPG object """
target = []
for i, obs in enumerate(obs_all_agents):
target_actions = [self.maddpg_agent[i].target_act(obs[i,:], batch = batch) for i in range(agent_no)]
target_actions = torch.stack(target_actions)
target.append(target_actions)
return target
def target_act_batch(self,agent_no, obs_all_agents, noise=0.0 , batch = True):
"""get target network actions from all the agents in the MADDPG object """
target_actions = [self.maddpg_agent[i].target_act(obs_all_agents[:,i,:] , batch = batch) for i in range(agent_no)]
target_actions = torch.stack(target_actions)
return target_actions
def act_with_agent(self,agent_no, agent_id ,obs_all_agents, noise=0.0 , batch = True):
"""get target network actions from all the agents in the MADDPG object """
actions = [self.maddpg_agent[i].actor(obs_all_agents[:,i,:] ,batch = batch) if i == agent_id \
else self.maddpg_agent[i].actor(obs_all_agents[:,i,:] , batch = batch).detach() for i in range(agent_no)]
actions = torch.stack(actions)
return actions
def update(self, samples, agent_number , update_actor = True):
"""update the critics and actors of all the agents """
# need to transpose each element of the samples
# to flip obs[parallel_agent][agent_number] to
# obs[agent_number][parallel_agent]
obs, obs_full, action, reward, next_obs, next_obs_full, done = samples
samples = (np.array(obs), obs_full, action, reward, np.array(next_obs), next_obs_full, done)
batch_size = np.shape(obs_full)[0]
Batch_use = True if batch_size >1 else False
action_size = self.num_agents * self.action_vector
obs, obs_full, action, reward, next_obs, next_obs_full, done = map(transpose_to_tensorAsitis, samples)
obs_full = torch.stack(obs_full).to(device)
next_obs_full = torch.stack(next_obs_full).to(device)
agent = self.maddpg_agent[agent_number]
agent.critic_optimizer.zero_grad()
# critic loss = batch mean of (y- Q(s,a) from target network)^2
#y = reward of this timestep + discount * Q(st+1,at+1) from target network
# !crictal logic error have to change her to the agents observation only
next_obs = torch.stack(next_obs).to(device)
obs = torch.stack(obs).to(device)
target_actions = self.target_act_batch(2,next_obs)
target_actions = target_actions.to(device)
#batch size and action size
#target_critic_input = torch.cat((next_obs_full.view(-1,batch_size).t(),target_actions.view(-1,action_size)), dim=1).to(device)
Current_Agent_actions_target,other_agent_Action_target = giveCurrentAgentsAction(target_actions , agent_number , Tuples = False, batch = Batch_use)
#print(Current_Agent_actions_target.size(),other_agent_Action_target.size() , other_agent_Action_target.view(self.action_vector,-1).size())
#print(next_obs_full.view(-1,batch_size).t().size() , next_obs_full.size() , next_obs_full)
with torch.no_grad():
critic_state = torch.cat((next_obs_full.view(-1,batch_size).t(), other_agent_Action_target) , dim=1)
q_next=agent.target_critic.critic_forward(critic_state,Current_Agent_actions_target.view(-1,self.action_vector)).to(device)
#indices = torch.tensor([1])
reward = torch.stack(reward).to(device)
done = torch.stack(done).to(device)
y = reward[:,agent_number].view(-1, 1) + self.discount_factor * q_next * (1 - done[:,agent_number].view(-1, 1)).to(device)
action = torch.stack(action).to(device)
Current_Agent_actions,other_agent_Action = giveCurrentAgentsAction(action , agent_number , batch = Batch_use)
#print(action,Current_Agent_actions.size(),other_agent_Action.size())
critic_input = torch.cat((obs_full.view(-1,batch_size).t(), other_agent_Action), dim=1).to(device)
q = agent.critic.critic_forward(critic_input, Current_Agent_actions.view(-1,self.action_vector)).to(device)
huber_loss = torch.nn.SmoothL1Loss()
mse_loss = F.mse_loss
critic_loss = huber_loss(q, y.detach())
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(agent.critic.parameters(), 0.5)
agent.critic_optimizer.step()
#update actor network using policy gradient
if(update_actor):
agent.actor_optimizer.zero_grad()
# make input to agent
# detach the other agents to save computation
# saves some time for computing derivative
# q_input = [ self.maddpg_agent[i].actor(ob) if i == agent_number \
# else self.maddpg_agent[i].actor(ob).detach()
# for i, ob in enumerate(obs) ]
q_input = self.act_with_agent(2,agent_number,obs)
q_input = q_input.to(device)
Current_Agent_Qin,other_agent_Qin = giveCurrentAgentsAction(q_input , agent_number , batch = Batch_use , Tuples = False)
# combine all the actions and observations for input to critic
# many of the obs are redundant, and obs[1] contains all useful information already
q_input2 = torch.cat((obs_full.view(-1,batch_size).t(), other_agent_Qin), dim=1).to(device)
# get the policy gradient
actor_loss = -agent.critic.critic_forward(q_input2, Current_Agent_Qin.view(-1,self.action_vector)).mean()
actor_loss.backward()
torch.nn.utils.clip_grad_norm_(agent.actor.parameters(),0.5)
agent.actor_optimizer.step()
# al = actor_loss.cpu().detach().item()
# cl = critic_loss.cpu().detach().item()
# logger.add_scalars('agent%i/losses' % agent_number,
# {'critic loss': cl,
# 'actor_loss': al},
# self.iter)
def update_targets(self):
"""soft update targets"""
self.iter += 1
for ddpg_agent in self.maddpg_agent:
soft_update(ddpg_agent.target_actor, ddpg_agent.actor, self.tau)
soft_update(ddpg_agent.target_critic, ddpg_agent.critic, self.tau)