Repository navigation
Expand file tree
/
Copy pathdp.py
More file actions
163 lines (152 loc) · 5.61 KB
/
Copy pathdp.py
File metadata and controls
163 lines (152 loc) · 5.61 KB
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
from env import GridWorld
import numpy as np
def value_iteration(env: GridWorld, max_iter=20):
env.reset()
# initialize policy
env.pi = np.ones((env.size, env.size, env.action_number), dtype=np.float32) / env.action_number
i = 0
while True:
tmp_pi = np.zeros((env.size, env.size, env.action_number), dtype=np.float32)
tmp_state_value = np.zeros((env.size, env.size), dtype=np.float32)
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
best_a = 0
best_q = -float('inf')
for a in range(env.action_number):
s2, r = env.take_action(s, a)
q_value = r + env.gamma * env.state_value[s2]
if q_value >= best_q:
best_q = q_value
best_a = a
# policy update
tmp_pi[s] = np.eye(env.action_number)[best_a]
# value update
tmp_state_value[s] = best_q
delta = np.max(np.abs(env.state_value - tmp_state_value))
env.pi = tmp_pi
env.state_value = tmp_state_value
if delta < 1e-4 or i >= max_iter:
break
else:
i += 1
print("Final Results after Value Iteration")
print(f"state_value :\n {env.state_value}")
def policy_iteration(env: GridWorld, max_iter=250):
env.reset()
# initialize policy
env.pi = np.ones((env.size, env.size, env.action_number), dtype=np.float32) / env.action_number
i = 0
while True:
# policy evaluation
while True:
# print(f" Evaluation Iteration: {j}")
max_delta = 0
tmp_state_value = env.state_value.copy()
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
v =0
for a in range(env.action_number):
s2, r = env.take_action(s, a)
v += env.pi[s][a] * (r + env.gamma * env.state_value[s2])
delta = np.abs(tmp_state_value[s] - v)
if delta > max_delta:
max_delta = delta
tmp_state_value[s] = v
if max_delta < 10: # max_delta 越小,收敛越慢
env.state_value = tmp_state_value
break
# policy improvement
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
best_a = 0
best_q = -float('inf')
for a in range(env.action_number):
s2, r = env.take_action(s, a)
q_value = r + env.gamma * env.state_value[s2]
if q_value > best_q:
best_q = q_value
best_a = a
# policy update
env.pi[s] = np.eye(env.action_number)[best_a]
if i >= max_iter:
break
else:
i += 1
print("Final Results after Policy Iteration")
print(f"state_value :\n {env.state_value}")
def truncated_policy_iteration(env: GridWorld, max_iter=250, max_eval_iter=10):
# initialize policy
env.pi = np.ones((env.size, env.size, env.action_number), dtype=np.float32) / env.action_number
i = 0
while True:
# policy evaluation
j = 0
while True:
# print(f" Evaluation Iteration: {j}")
max_delta = 0
tmp_state_value = env.state_value.copy()
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
v =0
for a in range(env.action_number):
s2, r = env.take_action(s, a)
v += env.pi[s][a] * (r + env.gamma * env.state_value[s2])
delta = np.abs(tmp_state_value[s] - v)
if delta > max_delta:
max_delta = delta
tmp_state_value[s] = v
if max_delta < 1e-4 or j >= max_eval_iter:
env.state_value = tmp_state_value
break
else:
j += 1
# policy improvement
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
best_a = 0
best_q = -float('inf')
for a in range(env.action_number):
s2, r = env.take_action(s, a)
q_value = r + env.gamma * env.state_value[s2]
if q_value > best_q:
best_q = q_value
best_a = a
# policy update
env.pi[s] = np.eye(env.action_number)[best_a]
if i >= max_iter:
break
else:
i += 1
print("Final Results after Truncated Policy Iteration")
print(f"state_value :\n {env.state_value}")
if __name__ == "__main__":
size = 2
forbidden_states = [(0, 1)]
target_state = (1, 1)
gamma = 0.9
forbidden_penalty = -1
boundary_penalty = -1
target_reward = 1
env1 = GridWorld(
size=size,
forbidden_states=forbidden_states,
target_state=target_state,
gamma=gamma,
forbidden_penalty=forbidden_penalty,
boundary_penalty=boundary_penalty,
target_reward=target_reward
)
value_iteration(env1)
size = 5
forbidden_states = [(1,1),(1, 2),(2,2),(3,1),(3,3),(4,1)]
target_state = (3,2)
gamma = 0.9
forbidden_penalty = -10
boundary_penalty = -1
target_reward = 1
env2 = GridWorld(
size=size,
forbidden_states=forbidden_states,
target_state=target_state,
gamma=gamma,
forbidden_penalty=forbidden_penalty,
boundary_penalty=boundary_penalty,
target_reward=target_reward
)
policy_iteration(env2)
truncated_policy_iteration(env2)