-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpingpong_konda.py
196 lines (147 loc) · 6.39 KB
/
pingpong_konda.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
import numpy as np
import tensorflow as tf
import keras.backend as K
from keras.layers import Dense, Activation, Input, Conv2D, Flatten
from keras.models import Model, load_model
from keras.optimizers import Adam,RMSprop
from keras.regularizers import l2
import time
class Actor():
def __init__(self, ALPHA, n_actions =4,
last_size=16, input_dims = 8):
self.lr = ALPHA
self.input_dims = input_dims
self.h1_dims = last_size
self.n_actions = n_actions
self.state_memory = []
self.action_memory = []
self.Q_memory = []
# Action values to send to gym environment to move paddle up/down
self.UP_ACTION = 2
self.DOWN_ACTION = 3
self.actor, self.phi, self.policy = self.build_polic_network()
self.actions_space = [i for i in range(n_actions)]
def build_polic_network(self):
input = Input(shape=(*self.input_dims,))
advantages = Input(shape=[1])
conv1 = Conv2D(filters=32, kernel_size=8, strides=4, activation='relu',
data_format='channels_first')(input)
conv2 = Conv2D(filters=64, kernel_size=4, strides=2, activation='relu',
data_format='channels_first')(conv1)
conv3 = Conv2D(filters=64, kernel_size=3, strides=1, activation='relu',
data_format='channels_first')(conv2)
flat = Flatten()(conv3)
dense1 = Dense(self.h1_dims, activation='relu')(flat)
probs = Dense(1, activation='sigmoid')(dense1)
#Loss funciton implimenting Cross Entropy
def custum_loss(y_true,y_pred):
#Clipping to ignore getting 0 and 1 has input from softmax layer
out = K.clip(y_pred, 1e-8,1-1e-8)
log_lik = y_true*K.log(out)
return K.sum(-log_lik*advantages)
#Generate the model with proper inputs and outputs
actor = Model(inputs=[input,advantages], outputs=[probs])
actor.compile(optimizer=RMSprop(lr=self.lr), loss=custum_loss)
actor.summary()
phi = Model(inputs=[input], outputs=[dense1])
predict = Model(input=[input], output=[probs])
# phi.compile(optimizer=Adam(lr=self.lr), loss="custum_loss(Q,y_true,y_pred)")
return actor, phi, predict
def store_transition(self, observation, action, Q):
self.action_memory.append(action)
self.state_memory.append(observation)
self.Q_memory.append(Q)
def learn(self):
state_memory = np.array(self.state_memory)
action_memory = np.array(self.action_memory)
Q_memory = np.array(self.Q_memory)
Q_memory = Q_memory - np.mean(Q_memory)
Q_memory/=np.std(Q_memory)
actions = np.zeros([len(action_memory), self.n_actions])
actions[np.arange(len(action_memory)),action_memory] = 1
cost = self.actor.train_on_batch([state_memory, Q_memory], actions)
self.state_memory = []
self.action_memory = []
self.Q_memory = []
# return cost
def choose_action(self, observation):
state = observation[np.newaxis, :]
#Get the probability for each action
probability = self.policy.predict(state)[0]
#Get the action by sampling from the given probability
if np.random.uniform() < probability:
action = self.UP_ACTION
else:
action = self.DOWN_ACTION
return action
return action
def save_model(self,name):
self.actor.save(name)
def load_model(self,name):
self.actor = load_model(name)
class Critic():
def __init__(self, ALPHA, Gamma = 0.99, n_actions =4,
input_dims = 8,decay=0.01):
self.gamma = Gamma
self.lr = ALPHA
self.decay = decay
self.input_dims = input_dims
self.n_actions = n_actions
self.actions_space = [i for i in range(n_actions)]
self.phi_memory = []
self.reward_memory = []
self.critic = self.build_polic_network()
def build_polic_network(self):
input = Input(shape=(self.input_dims,))
Qvalue = Dense(1, activation='linear',use_bias=False)(input)
critic = Model(input = [input], output = [Qvalue])
critic.compile(optimizer=Adam(lr=self.lr), loss='mean_squared_error')
critic.summary()
return critic
def store_transition(self, phi, reward):
self.phi_memory.append(phi)
self.reward_memory.append(reward)
def getPhi(self, actor_hidden_out, probs, action):
Q_input = np.zeros(self.input_dims)
probs = [-prob for prob in probs]
probs[action] = 1 + probs[action]
output_size = len(actor_hidden_out)
#Geting input for critic linear approximation as given
# in Policy Gradient Methods for
#Reinforcement Learning with FunctionApproximation
#by Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
for i in range(self.n_actions):
Q_input = Q_input + np.pad(actor_hidden_out,\
(i*output_size,(self.n_actions-1-i)*output_size)\
, 'constant') * probs[i]
return Q_input
def learn(self):
phi_memory = np.array(self.phi_memory)
reward_memory = np.array(self.reward_memory)
#Get future reward for each state in an episode
G = np.array(np.zeros_like(reward_memory))[:,np.newaxis]
for r in range(len(reward_memory)):
G_sum = 0
discount = 1
for k in range(r, len(reward_memory)):
G_sum +=reward_memory[k]*discount
discount *= self.gamma
if reward_memory[k] != 0:
# Don't count rewards from subsequent rounds
break
G[r][0] = G_sum
#Here we are considering G has target funtion
#But need to replace G for the current action performed
#Rest should be equal to their Q value so the loss
#for other Q for other actions must be zero
# target = self.critic.predict(phi_memory)
# target[np.arange(len(action_memory)),action_memory] = G
cost = self.critic.train_on_batch(phi_memory, G)
# print(self.critic.get_weights())
self.phi_memory = []
self.reward_memory = []
# return cost
def save_model(self,name):
self.critic.save(name)
def load_model(self,name):
self.critic = load_model(name)