-
Notifications
You must be signed in to change notification settings - Fork 3
/
rnnl.py
162 lines (141 loc) · 6.76 KB
/
rnnl.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
# -*- coding: utf-8 -*-
"""
@author: Jin Zhu
"""
import tensorflow as tf
from tensorflow.keras import layers
import numpy as np
from sklearn.metrics.pairwise import euclidean_distances
class RatioRKHSLearner(tf.Module):
def __init__(self, hidden_node, truncate=20):
super(RatioRKHSLearner, self).__init__()
self.hidden_layer = layers.Dense(hidden_node, activation='relu', dtype='float64')
self.output_layer = layers.Dense(1, dtype='float64')
self.truncate = truncate
def __call__(self, x):
x = self.hidden_layer(x)
x = self.output_layer(x)
return x
def get_ratio_prediction(self, state, normalize=True):
if state.shape == ():
state = state.reshape(1, 1)
elif state.shape == (1, ):
state = state.reshape(1, state.size)
else:
pass
ratio = self.__call__(state)
ratio = ratio.numpy()
ratio = ratio.flatten()
## truncate:
ratio_min = 1 / self.truncate
ratio_max = self.truncate
ratio = np.clip(ratio, a_min=ratio_min, a_max=ratio_max)
if state.shape[0] > 1:
if normalize:
ratio /= np.mean(ratio)
return ratio
def get_r_prediction(self, state, normalize=True):
return self.get_ratio_prediction(state, normalize)
def tf_gaussian_kernel(x, width):
kernel_value = np.divide(np.exp(-x), 2.0*np.square(width))
return kernel_value
def loss(model, s0, state, next_state, policy_ratio, width, gamma, batch_size):
loss_term1_value = loss_term_1(model, s0, state, next_state, policy_ratio, width, gamma, batch_size)
loss_term2_value = loss_term2(model, state, next_state, policy_ratio, width, gamma, batch_size)
loss_term = loss_term1_value + loss_term2_value
return loss_term
def term1_gap(s0, state, next_state, policy_ratio, width, gamma, batch_size):
gap_vec = np.zeros(batch_size)
for i in range(batch_size):
s0_state_dist = s0 - state[i, :]
s0_state_dist = np.sqrt(np.sum(np.square(s0_state_dist), axis=1))
kern_s0_s = tf_gaussian_kernel(s0_state_dist, width)
s0_next_state_dist = s0 - next_state[i, :]
s0_next_state_dist = np.sqrt(np.sum(np.square(s0_next_state_dist), axis=1))
kern_s0_ns = tf_gaussian_kernel(s0_next_state_dist, width)
gap = np.mean(kern_s0_s - gamma * policy_ratio[i] * kern_s0_ns)
gap_vec[i] = gap
return gap_vec
def loss_term_1(model, s0, state, next_state, policy_ratio, width, gamma, batch_size):
weight_vec = term1_gap(s0, state, next_state, policy_ratio, width, gamma, batch_size)
state_fit = model(state)
term_1plus2_value = -2.0 * (1.0 - gamma) * tf.reduce_mean(state_fit * weight_vec)
return term_1plus2_value
def term2_gap(state, next_state, policy_ratio, width, gamma):
cross_ratio = np.square(gamma) * policy_ratio.reshape(-1, 1) * tf.transpose(policy_ratio)
cross_kernel_ns = tf_gaussian_kernel(euclidean_distances(next_state), width)
cross_kernel_s = tf_gaussian_kernel(euclidean_distances(state), width)
cross_kernel_s_ns = tf_gaussian_kernel(euclidean_distances(state, next_state), width)
cross_kernel_s_ns = cross_kernel_s_ns * policy_ratio
gap_mat = cross_ratio * cross_kernel_ns + cross_kernel_s - 2.0 * gamma * cross_kernel_s_ns
return gap_mat
def loss_term2(model, state, next_state, policy_ratio, width, gamma, batch_size):
state_fit = model(state)
cross_pred_mat = tf.reshape(state_fit, [-1, 1]) * tf.transpose(state_fit)
gap_mat = term2_gap(state, next_state, policy_ratio, width, gamma)
term3_value = tf.reduce_mean(cross_pred_mat * gap_mat)
return term3_value
def gradient(model, s0, state, next_state, policy_ratio, width, gamma, batch_size):
with tf.GradientTape() as tape:
loss_value = loss(model, s0, state, next_state, policy_ratio, width, gamma, batch_size)
return loss_value, tape.gradient(loss_value, model.trainable_variables)
def train_one_step(model, optimizer, s0, state, next_state, policy_ratio, width, gamma, batch_size):
loss_value, grads = gradient(model, s0, state, next_state, policy_ratio, width, gamma, batch_size)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
return loss_value
def compute_pm_ratio(state, action, mediator, policy, cplearner, use_mediator=True):
if use_mediator:
unique_action = np.unique(action)
ratio = np.zeros((action.shape[0], 1))
for action_value in unique_action:
action_value = np.array([action_value])
target_pa = np.apply_along_axis(policy, 1, state, action=action_value).flatten()
policy_action_tmp = np.repeat(action_value, action.shape[0]).reshape(-1, 1)
pm_ratio = cplearner.get_pm_ratio(state, policy_action_tmp, action, mediator)
ratio += (pm_ratio * target_pa).reshape(-1, 1)
ratio = ratio.flatten()
else:
estimate_pa = cplearner.get_pa_prediction(state, action)
target_pa = np.zeros([0.0])
pa_ratio = target_pa / estimate_pa
ratio = pa_ratio
return ratio
def compute_median_distance(state, max_obs=500):
if state.shape[0] <= max_obs:
median_state_dist = np.median(euclidean_distances(state, state))
else:
sub_index = np.random.choice(state.shape[0], max_obs)
state_part = state[sub_index, :]
median_state_dist = np.median(euclidean_distances(state_part, state_part))
return median_state_dist
def train(model, optimizer, dataset, policy, cplearner, gamma=0.9, batch_size=2048, epoch=1000, trace=True):
s0 = dataset['s0']
state = dataset['state']
action = dataset['action']
mediator = dataset['mediator']
policy_ratio = compute_pm_ratio(state, action, mediator, policy, cplearner)
next_state = dataset['next_state']
median_state_dist = compute_median_distance(state)
train_data_num = state.shape[0]
train_loss_results = []
batch_size = min(state.shape[0], batch_size)
if trace:
print("Start Training")
for i in range(1, 1+epoch):
np.random.seed(i)
idx = np.random.permutation(train_data_num)
state = state[idx, :]
next_state = next_state[idx, :]
policy_ratio = policy_ratio[idx]
for k in range(int(train_data_num/batch_size)):
s = state[k*batch_size:(k+1)*batch_size, :]
ns = next_state[k*batch_size:(k+1)*batch_size, :]
pr = policy_ratio[k*batch_size:(k+1)*batch_size]
loss_value = train_one_step(model, optimizer, s0, s, ns, pr, median_state_dist, gamma, batch_size)
# print(loss_value)
# print(loss_value.shape)
train_loss_results.append(loss_value)
if trace and (i % 5 == 0):
print("Epoch: {0}; Loss: {1}".format(i, loss_value))
pass
pass