TP1_Reinforcment_Learning_Py/main_tp1_ex1.py at master · AntoinePlumerault/TP1_Reinforcment_Learning_Py · GitHub

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
from mdp import MDP
import numpy as np
import matplotlib.pyplot as plt
import time

P_x_a = np.array([[[0.45, 0.55, 0.00],
		           [0.60, 0.40, 0.00],
		           [0.00, 1.00, 0.00]],
	              [[0.00, 1.00, 0.00],
	               [0.00, 0.10, 0.90],
	               [0.50, 0.10, 0.40]]])

R_x_a = np.array([[-0.4, -1.0, 2.0],
	              [ 0.0, -0.5, 0.0]])

env = MDP(P_x_a, R_x_a)

X = range(env.n_states)
A = range(env.n_actions)
gamma = 0.95

# ========= Policy evaluation ========= #

pi_opt = [1,1,0]  # Optimal policy

V_opt = np.array([0.0] * env.n_states)
T = 10000
for t in range(T):
	for x in X:
		next_state, reward = env.step(x, pi_opt[x])
		V_opt[x] = reward + gamma*sum(P_x_a[pi_opt[x], x, :] * V_opt[:]) # WARNING


# ========== Value iteration ========== #
t = time.time()

Delta_V_Value = []

V = np.array([0.0] * env.n_states)
term = False
while not term:
	TV = np.array([0.0] * env.n_states)
	for x in X:
		TV[x] = -float('inf')
		for a in A:
			next_state, reward = env.step(x, a)
			TVa = reward + gamma*sum(P_x_a[a, x, :] * V[:])
			if TVa > TV[x]:
				TV[x] = TVa

	#print(max(abs(TV - V)))
	if max(abs(TV - V)) < 0.01:
		term = True

	Delta_V_Value.append(max(abs(V_opt - V)))

	V = TV

pi = [A[0]] * env.n_states
for x in X:
	best_a = A[0]
	best_Va = -float('inf')
	for a in A:
		next_state, reward = env.step(x, a)
		Va = reward + gamma*sum(P_x_a[a, x, :] * V[:])
		if Va > best_Va:
			best_a = a
			best_Va = Va
	pi[x] = best_a

elapsed = time.time() - t

plt.plot(Delta_V_Value)

print('elapsed Value Iteration', elapsed)

# ========= Policy iteration ========== #

t = time.time()

Delta_V_Policy = []

pi = [A[0]] * env.n_states
V = np.array([0.0] * env.n_states)
term = False
while (not term):
	V_pi = np.array([0.0] * env.n_states)
	for x in X:
		next_state, reward = env.step(x, pi[x])
		V_pi[x] = reward + gamma*sum(P_x_a[pi[x], x, :] * V[:])

	for x in X:
		best_a = A[0]
		best_Va = -float('inf')
		for a in A:
			next_state, reward = env.step(x, a)
			Va = reward + gamma*sum(P_x_a[a, x, :] * V[:])
			if Va > best_Va:
				best_a = a
				best_Va = Va
		pi[x] = best_a

	Delta_V_Policy.append(max(abs(V_opt - V)))

	if max(abs(V_pi - V)) < 0.01:
		term = True

	V = V_pi

elapsed = time.time() - t

plt.plot(Delta_V_Policy)

print(len(Delta_V_Policy),len(Delta_V_Value))
print('elapsed Policy Iteration', elapsed)

# ====================================== #

plt.show()

print(2*0.01 * gamma / (1.0-gamma))
print(pi)
print(V)
print(V_opt)