-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathSOTO-ALF-CLDE.py
executable file
·174 lines (145 loc) · 6.66 KB
/
SOTO-ALF-CLDE.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
import numpy as np
from keras.utils import to_categorical
import copy
from common.utils import eligibility_traces, default_config, make_env, str2bool, get_more_obs_com, discount_rewards
from common.ppo_independant import PPOPolicyNetwork, ValueNetwork
from collections import deque
render = False
normalize_inputs = True
config = default_config()
LAMBDA = float(config['agent']['lambda'])
lr_actor = float(config['agent']['lr_actor'])
alpha_fairness = float(config['agent']['alpha_fairness'])
twophase_proportion = float(config['agent']['twophase_proportion'])
env = make_env(config, normalize_inputs)
env.toggle_compute_neighbors()
n_agent = env.n_agent
T = env.T
GAMMA = env.GAMMA
n_episode = env.n_episode
max_steps = env.max_steps
n_actions = env.n_actions
i_episode = 0
gPi = []
Pi = []
gV = []
V = []
more_obs_size=env.neighbors_size+1
more_obs_size2 = n_agent
for i in range(n_agent):
gPi.append(PPOPolicyNetwork(num_features=env.input_size, num_actions=n_actions, layer_size=256, epsilon=0.1,
learning_rate=lr_actor))
Pi.append(PPOPolicyNetwork(num_features=env.input_size+more_obs_size+more_obs_size2+n_actions, num_actions=n_actions, layer_size=64, epsilon=0.1,
learning_rate=lr_actor))
gV.append(ValueNetwork(num_features=env.input_size, hidden_size=256, learning_rate=0.001))
V.append(ValueNetwork(num_features=env.input_size+more_obs_size+more_obs_size2+n_actions, hidden_size=256, learning_rate=0.001))
memory_ep_rewards = [deque() for _ in range(n_agent)]
average_jpi = np.zeros(n_agent)
while i_episode < n_episode:
beta = max(1 - float(i_episode) / (twophase_proportion * float(n_episode)), 0.0)
i_episode += 1
memory_ep_rewards = [deque() for _ in range(n_agent)]
average_jpi = np.zeros(n_agent)
avg = [0.] * n_agent
ep_actions = [[] for _ in range(n_agent)]
ep_rewards = [[] for _ in range(n_agent)]
ep_states = [[] for _ in range(n_agent)]
greedy = np.zeros(n_agent).astype(bool)
for i in range(n_agent):
greedyc = np.random.rand() <= beta
greedy[i] = greedyc
score = 0
steps = 0
su = [0.] * n_agent
su = np.array(su)
obs = env.reset()
neighbors = env.neighbors()
done = False
while steps < max_steps and not done:
steps += 1
action = []
for i in range(n_agent):
h = copy.deepcopy(obs[i])
if not greedy[i]:
more_obs = gPi[i].get_dist(np.array([h]))[0]
h.extend(more_obs)
more_obs = get_more_obs_com(True, neighbors, average_jpi, i, more_obs_size)
h.extend(more_obs)
more_obs = average_jpi
more_obs = (more_obs - np.mean(more_obs)) / (np.std(more_obs) + 0.0000000001)
h.extend(more_obs)
p = Pi[i].get_dist(np.array([h]))[0]
else:
p = gPi[i].get_dist(np.array([h]))[0]
ep_states[i].append(h)
action.append(np.random.choice(range(n_actions), p=p))
ep_actions[i].append(to_categorical(action[i], n_actions))
obs, rewards, done = env.step(action)
neighbors = env.neighbors()
su += np.array(rewards)
score += sum(rewards)
for i in range(n_agent):
ep_rewards[i].append(rewards[i])
memory_ep_rewards[i].append(rewards[i])
average_jpi[i] += rewards[i]
if len(memory_ep_rewards[i]) > max_steps * 5:
average_jpi[i] -= memory_ep_rewards[i].popleft()
if steps % T == 0:
all_ep_advantages = []
all_ep_advantages_unnormed = []
for i in range(n_agent):
ep_actions[i] = np.array(ep_actions[i])
ep_rewards[i] = np.array(ep_rewards[i], dtype=np.float_)
ep_states[i] = np.array(ep_states[i])
if LAMBDA < -0.1:
targets = discount_rewards(ep_rewards[i], GAMMA)
if not greedy[i]:
V[i].update(ep_states[i], targets)
vs = V[i].get(ep_states[i])
else:
gV[i].update(ep_states[i], targets)
vs = gV[i].get(ep_states[i])
else:
next_s = copy.deepcopy(obs[i])
if not greedy[i]:
vs = V[i].get(ep_states[i])
more_obs = gPi[i].get_dist(np.array([obs[i]]))[0]
next_s.extend(more_obs)
more_obs = get_more_obs_com(True, neighbors, average_jpi, i, more_obs_size)
next_s.extend(more_obs)
more_obs = average_jpi
more_obs = (more_obs - np.mean(more_obs)) / (np.std(more_obs) + 0.0000000001)
next_s.extend(more_obs)
targets = eligibility_traces(ep_rewards[i], vs, V[i].get([next_s]), GAMMA, LAMBDA)
V[i].update(ep_states[i], targets)
else:
vs = gV[i].get(ep_states[i])
targets = eligibility_traces(ep_rewards[i], vs, gV[i].get([next_s]), GAMMA, LAMBDA)
gV[i].update(ep_states[i], targets)
ep_advantages = targets - vs
all_ep_advantages_unnormed.append(ep_advantages)
ep_advantages = (ep_advantages - np.mean(ep_advantages)) / (np.std(ep_advantages) + 0.0000000001)
all_ep_advantages.append(ep_advantages)
all_ep_advantages = np.array(all_ep_advantages)
all_ep_advantages_unnormed = np.array(all_ep_advantages_unnormed)
all_ep_advantages_unnormed = ((average_jpi+0.0000000001)**-alpha_fairness) @ all_ep_advantages_unnormed
all_ep_advantages_normed = (all_ep_advantages_unnormed - np.mean(all_ep_advantages_unnormed)) / (np.std(all_ep_advantages_unnormed) + 0.0000000001)
for i in range(n_agent):
if not greedy[i]:
Pi[i].update(ep_states[i], ep_actions[i], all_ep_advantages_normed)
else:
gPi[i].update(ep_states[i], ep_actions[i], all_ep_advantages[i])
ep_actions = [[] for _ in range(n_agent)]
ep_rewards = [[] for _ in range(n_agent)]
ep_states = [[] for _ in range(n_agent)]
greedy=np.zeros(n_agent).astype(bool)
for i in range(n_agent):
greedyc = np.random.rand() <= beta
greedy[i] = greedyc
if render:
env.render()
print(i_episode)
print(score / max_steps)
print(su)
print(env.rinfo.flatten())
env.end_episode()