-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathaveraged_dqn.py
178 lines (157 loc) · 6.29 KB
/
averaged_dqn.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
import argparse
import sys
import copy
from collections import deque
import gym
import numpy as np
import chainer.links as L
import chainer.functions as F
from chainer import Chain, optimizers, Variable, serializers
parser = argparse.ArgumentParser(description='CVAE')
parser.add_argument('--env', '-e', type=str, default="CartPole-v0",
help='Open AI environment')
parser.add_argument('--Episode', '-episode', default=400, type=int,
help='number of episode to learn')
parser.add_argument('--K', '-k', type=int, default=5,
help='average size')
args = parser.parse_args()
class Network(Chain):
def __init__(self, n_in, n_out):
super(Network, self).__init__(
L1=L.Linear(n_in, 100),
L2=L.Linear(100, 200),
L3=L.Linear(200, 100),
L4=L.Linear(100, 100),
q_value=L.Linear(100, n_out, initialW=np.zeros((n_out, 100), dtype=np.float32))
)
def q_func(self, in_layer):
layer1 = F.leaky_relu(self.L1(in_layer))
layer2 = F.leaky_relu(self.L2(layer1))
layer3 = F.leaky_relu(self.L3(layer2))
layer4 = F.leaky_relu(self.L4(layer3))
return self.q_value(layer4)
class Agent():
def __init__(self, n_state, n_action, seed):
np.random.seed(seed)
sys.setrecursionlimit(10000)
self.n_action = n_action
self.model = Network(n_state, n_action)
self.optimizer = optimizers.Adam()
self.optimizer.setup(self.model)
self.memory = deque()
self.loss = 0
self.step = 0
self.train_freq = 10
self.target_update_freq = 20
self.gamma = 0.99
self.mem_size = 1000
self.replay_size = 100
self.epsilon = 0.05
def stock_experience(self, exp):
self.memory.append(exp)
if len(self.memory) > self.mem_size:
self.memory.popleft()
def forward(self, exp,target_model):
state = Variable(exp["state"])
state_dash = Variable(exp["state_dash"])
q_action = self.model.q_func(state)
tmp=0
for i in range(args.K):
tmp += target_model[i].q_func(state_dash)
tmp=tmp/args.K
tmp = list(map(np.max, tmp.data))
max_q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(copy.deepcopy(q_action.data), dtype=np.float32)
for i in range(self.replay_size):
target[i, exp["action"][i]] = exp["reward"][i] \
+ (self.gamma * max_q_dash[i]) * (not exp["ep_end"][i])
loss = F.mean_squared_error(q_action, Variable(target))
self.loss = loss.data
return loss
def action(self, state):
if np.random.rand() < self.epsilon:
return np.random.randint(0, self.n_action)
else:
state = Variable(state)
q_action = self.model.q_func(state)
q_action = q_action.data[0]
act = np.argmax(q_action)
return np.asarray(act, dtype=np.int8)
def experience_replay(self,target_model):
mem = np.random.permutation(np.array(self.memory))
perm = np.array([i for i in range(len(mem))])
tmp_loss=0
for start in perm[::self.replay_size]:
index = perm[start:start+self.replay_size]
replay = mem[index]
state = np.array([replay[i]["state"] \
for i in range(self.replay_size)], dtype=np.float32)
action = np.array([replay[i]["action"] \
for i in range(self.replay_size)], dtype=np.int8)
reward = np.array([replay[i]["reward"] \
for i in range(self.replay_size)], dtype=np.float32)
state_dash = np.array([replay[i]["state_dash"] \
for i in range(self.replay_size)], dtype=np.float32)
ep_end = np.array([replay[i]["ep_end"] \
for i in range(self.replay_size)], dtype=np.bool)
experience = {"state":state, "action":action, \
"reward":reward, "state_dash":state_dash, "ep_end":ep_end}
self.model.zerograds()
loss = self.forward(experience,target_model)
tmp_loss+=loss.data
loss.backward()
self.optimizer.update()
return tmp_loss/self.replay_size
def train(self,target_model):
loss=-1
if len(self.memory) >= self.mem_size:
if self.step % self.train_freq == 0:
loss=self.experience_replay(target_model)
target_model.popleft()
target_model.append(copy.deepcopy(agent.model))
self.step += 1
return [target_model,loss]
def save_model(self, model_dir):
serializers.save_npz(model_dir + "model.npz", self.model)
def load_model(self, model_dir):
serializers.load_npz(model_dir + "model.npz", self.model)
self.target_model = copy.deepcopy(self.model)
env = gym.make(args.env)
n_state = env.observation_space.shape[0]
n_action = env.action_space.n
seed = 114514
agent = Agent(n_state, n_action, seed)
action_list = [i for i in range(0, n_action)]
target_model=deque()
for i in range(args.K):
target_model.append(copy.deepcopy(agent.model))
mat=np.zeros([args.Episode,3])
for _episode in range(args.Episode):
observation = env.reset()
total_reward=0
loss=0
count=0
for _times in range(200):#2000
env.render()
[target_model,tmp_loss]=agent.train(target_model)
if tmp_loss>=0:
loss+=tmp_loss
count+=1
state = observation.astype(np.float32).reshape((1, n_state))
action = action_list[agent.action(state)]
observation, reward, ep_end, _ = env.step(action)
state_dash = observation.astype(np.float32).reshape((1, n_state))
experience = {"state":state, "action":action, \
"reward":reward, "state_dash":state_dash, "ep_end":ep_end}
agent.stock_experience(experience)
total_reward+=reward
if ep_end:
if count==0:
ans=-1
else:
ans=loss/count
mat[_episode,:]=[_episode,ans,total_reward]
print(_episode,ans,total_reward)
break
file_name='./'+"Episode_"+str(args.Episode)+'_K_'+str(args.K)+'.summary_csv'
np.savetxt(file_name,mat,delimiter=',',header="Episode,Error,Reward")