-
Notifications
You must be signed in to change notification settings - Fork 0
/
paddpg.py
129 lines (99 loc) · 5.52 KB
/
paddpg.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
import copy
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from actor_critic import Actor, Critic
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class PADDPG(object):
def __init__(self, state_dim, action_dim, max_action, min_action, agents, exploration="EG", discount=0.99, tau=1e-4):
self.actor = Actor(state_dim, action_dim).to(device)
self.actor_target = copy.deepcopy(self.actor)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=1e-3)
self.critic = Critic(state_dim, action_dim).to(device)
self.critic_target = copy.deepcopy(self.critic)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=1e-3)
self.discount = discount
self.tau = tau
self.max_p = torch.FloatTensor(max_action).to(device)
self.min_p = torch.FloatTensor(min_action).to(device)
self.rng = (self.max_p - self.min_p).detach()
self.exploration = exploration
self.agents = agents
def invert_gradient(self, delta_a, current_a):
index = delta_a > 0
delta_a[index] *= (index.float() * (self.max_p - current_a) / self.rng)[index]
delta_a[~index] *= ((~index).float() * (current_a - self.min_p) / self.rng)[~index]
return delta_a
def select_action(self, state):
state = torch.FloatTensor(state.reshape(1, -1)).to(device)
p = self.actor(state)
np_max = self.max_p.cpu().data.numpy()
np_min = self.min_p.cpu().data.numpy()
return np.clip(p.cpu().data.numpy().flatten(), np_min, np_max)
def train(self, replay_buffer, batch_size=64):
if self.agents >= 2:
state, action, next_state, reward, ex_reward, n_step, ex_n_step, not_done, o_state,o_next_state,o_int_reward = replay_buffer.sample2(batch_size)
else:
state, action, next_state, reward, ex_reward, n_step, ex_n_step, not_done = replay_buffer.sample(batch_size)
################################# critic update #################################
target_Q = self.critic_target(next_state, self.actor_target(next_state))
# 探索法別にtargetQの計算を分ける
if self.exploration == "EG":
target_Q = reward + ((1-not_done) * self.discount * target_Q).detach()
else:
target_Q = (reward + ex_reward) + ((1-not_done) * self.discount * target_Q).detach()
current_Q = self.critic(state, action)
beta = 0.2
mixed_q = beta * n_step + (1 - beta) * target_Q
# mixed_q = beta * (n_step + ex_n_step) + (1 - beta) * target_Q
critic_loss = F.mse_loss(current_Q, mixed_q)
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 10)
self.critic_optimizer.step()
if self.agents >= 2:
target_Q = self.critic_target(o_next_state, self.actor_target(o_next_state))
target_Q = (reward+o_int_reward) + ((1-not_done) * self.discount * target_Q).detach()
current_Q = self.critic(o_state, self.actor(o_state))
beta = 0.2
#mixed_q = beta * n_step + (1 - beta) * target_Q
# mixed_q = beta * (n_step + ex_n_step) + (1 - beta) * target_Q
critic_loss = F.mse_loss(current_Q, target_Q)
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 10)
self.critic_optimizer.step()
################################# actor update #################################
current_a = Variable(self.actor(state))
current_a.requires_grad = True
actor_loss = self.critic(state, current_a).mean()
self.critic.zero_grad()
actor_loss.backward()
delta_a = copy.deepcopy(current_a.grad.data)
delta_a = self.invert_gradient(delta_a, current_a)
current_a = self.actor(Variable(state))
out = -torch.mul(delta_a, current_a)
self.actor.zero_grad()
out.backward(torch.ones(out.shape).to(device))
torch.nn.utils.clip_grad_norm_(self.actor.parameters(), 10)
self.actor_optimizer.step()
for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
# itemでpythonのint型として返す
return [current_Q.mean().item(), mixed_q.mean().item(), critic_loss.item()]
def save(self, filename):
torch.save(self.critic.state_dict(), filename + "_critic")
torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer")
torch.save(self.actor.state_dict(), filename + "_actor")
torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer")
def load(self, filename):
self.critic.load_state_dict(torch.load(filename + "_critic"))
self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer"))
self.critic_target = copy.deepcopy(self.critic)
self.actor.load_state_dict(torch.load(filename + "_actor"))
self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
self.actor_target = copy.deepcopy(self.actor)