Repository navigation
Expand file tree
/
Copy pathmc.py
More file actions
158 lines (127 loc) · 4.99 KB
/
Copy pathmc.py
File metadata and controls
158 lines (127 loc) · 4.99 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
import numpy as np
from env import GridWorld
def random_argmax(arr):
"""随机打破平均的 argmax"""
return np.random.choice(np.flatnonzero(arr == np.max(arr)))
def mc_explore_start(env: GridWorld, episodes=10000, max_episode_length=500):
env.reset()
# 为每个状态-动作对维护一个回报列表
returns = np.empty((env.size, env.size, env.action_number), dtype=object)
for i in range(env.size):
for j in range(env.size):
for a in range(env.action_number):
returns[i, j, a] = []
for ep in range(episodes):
s0 = (np.random.randint(0, env.size), np.random.randint(0, env.size))
a0 = np.random.randint(0, env.action_number)
trajectory = [(s0, a0)]
s_next, r = env.take_action(s0, a0)
rewards = [r]
while True:
action_probs = env.pi[s_next]
a_next = np.random.choice(np.arange(env.action_number), p=action_probs)
trajectory.append((s_next, a_next))
s, a = s_next, a_next
s_next, r = env.take_action(s, a)
rewards.append(r)
if len(trajectory) > max_episode_length:
break
T = len(trajectory)
G = 0
first_visit_dict = {}
for t in range(T):
s_t, a_t = trajectory[t]
if (s_t, a_t) not in first_visit_dict:
first_visit_dict[(s_t, a_t)] = t
for t in range(T-1, -1, -1):
G = env.gamma * G + rewards[t]
s_t, a_t = trajectory[t]
if (first_visit_dict[(s_t, a_t)] == t):
returns[s_t][a_t].append(G)
env.action_value[s_t][a_t] = np.mean(returns[s_t][a_t])
# 更新策略
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
best_a = random_argmax(env.action_value[s])
# 更新 pi 为确定性策略 (one-hot)
env.pi[s] = np.eye(env.action_number)[best_a]
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
env.pi_star[s] = random_argmax(env.action_value[s])
print("Pi after MC with Explore Start:")
print(env.pi_star)
def mc_epsilon_greedy(env: GridWorld, episodes=10000, max_episode_length=500, epsilon=0.1):
env.reset()
# 为每个状态-动作对维护一个回报列表
returns = np.empty((env.size, env.size, env.action_number), dtype=object)
for i in range(env.size):
for j in range(env.size):
for a in range(env.action_number):
returns[i, j, a] = []
for ep in range(episodes):
s0 = (np.random.randint(0, env.size), np.random.randint(0, env.size))
a0 = np.random.randint(0, env.action_number)
trajectory = [(s0, a0)]
s_next, r = env.take_action(s0, a0)
rewards = [r]
while True:
action_probs = env.pi[s_next]
a_next = np.random.choice(np.arange(env.action_number), p=action_probs)
trajectory.append((s_next, a_next))
s, a = s_next, a_next
s_next, r = env.take_action(s, a)
rewards.append(r)
if len(trajectory) > max_episode_length:
break
T = len(trajectory)
G = 0
first_visit_dict = {}
for t in range(T):
s_t, a_t = trajectory[t]
if (s_t, a_t) not in first_visit_dict:
first_visit_dict[(s_t, a_t)] = t
for t in range(T-1, -1, -1):
G = env.gamma * G + rewards[t]
s_t, a_t = trajectory[t]
if first_visit_dict[(s_t, a_t)] == t:
returns[s_t][a_t].append(G)
env.action_value[s_t][a_t] = np.mean(returns[s_t][a_t])
# 更新策略
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
best_a = random_argmax(env.action_value[s])
for a in range(env.action_number):
if a == best_a:
env.pi[s][a] = 1 - epsilon + (epsilon / env.action_number)
else:
env.pi[s][a] = epsilon / env.action_number
for s in [(i, j) for i in range(env.size) for j in range(env.size)]:
env.pi_star[s] = random_argmax(env.action_value[s])
print("Pi after MC with Epsilon-Greedy:")
print(env.pi_star)
if __name__ == "__main__":
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
env = 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
)
mc_explore_start(env)
mc_epsilon_greedy(env)
print("pi_star_GT:")
GT = [
[1,1,1,1,2],
[0,0,1,2,2],
[0,3,2,1,2],
[0,1,4,3,2],
[0,1,0,3,3],
]
GT = np.array(GT)
print(GT)