/
Agents.py
226 lines (179 loc) · 9.05 KB
/
Agents.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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import numpy as np
import tensorflow as tf
import logging
logging.basicConfig(filename='Agents.log',level=logging.DEBUG)
RANDOM_SEED = 8
np.random.seed(RANDOM_SEED)
tf.set_random_seed(RANDOM_SEED)
from enum import Enum, auto
class Critic_Variant(Enum):
INDEPENDENT = auto()
CENTRALIZED = auto()
CENTRALIZED_APPROX = auto()
class Agent(object):
def __init__(self, env, learning_rate=0.001, gamma = 0.95, agent_idx = 0):
self.sess = tf.Session()
self.env = env
self.n_actions = env.n_actions
self.n_features = env.n_features
self.learning_rate = learning_rate
self.gamma = gamma
self.agent_idx = agent_idx
self.log = [] # logs action probabilities
def choose_action(self, s):
action_probs = self.calc_action_probs(s)
action = np.random.choice(range(action_probs.shape[1]), p=action_probs.ravel()) # select action w.r.t the actions prob
self.log.append(action_probs[0,1])
return action
def learn_at_episode_end(self):
pass
def close(self):
self.sess.close()
tf.reset_default_graph()
def reset(self):
self.sess.run(tf.global_variables_initializer())
class Actor_Critic_Agent(Agent):
def __init__(self, env, learning_rate=0.001, n_units_actor = 20,
n_units_critic = 20, gamma = 0.95, agent_idx = 0,
critic_variant = Critic_Variant.INDEPENDENT, *args):
super().__init__(env, learning_rate, gamma, agent_idx)
self.actor = Actor(env, n_units_actor, learning_rate, agent_idx)
self.critic = Critic(env, n_units_critic, learning_rate, gamma, agent_idx,
critic_variant)
self.sess.run(tf.global_variables_initializer())
def learn(self, s, a, r, s_, done = False, *args):
if done:
pass
else:
td = self.critic.learn(self.sess,s,r,s_, *args)
self.actor.learn(self.sess,s,a,td)
def __str__(self):
return 'Actor_Critic_Agent_'+str(self.agent_idx)
def calc_action_probs(self, s):
return self.actor.calc_action_probs(self.sess,s)
def pass_agent_list(self, agent_list):
self.critic.pass_agent_list(agent_list)
def get_action_prob_variable(self):
return self.actor.actions_prob
def get_state_variable(self):
return self.actor.s
def get_policy_parameters(self):
return [self.actor.w_l1,self.actor.b_l1,self.actor.w_pi1,self.actor.b_pi1]
class Actor(object):
def __init__(self, env, n_units = 20, learning_rate=0.001, agent_idx = 0):
self.s = tf.placeholder(tf.float32, [1, env.n_features], "state")
self.a = tf.placeholder(tf.int32, None, "act")
self.td_error = tf.placeholder(tf.float32, None, "td_error") # TD_error
with tf.variable_scope('Actor'):
self.w_l1 = tf.Variable(tf.random_normal([env.n_features,n_units],stddev=0.1))
self.b_l1 = tf.Variable(tf.random_normal([n_units],stddev=0.1))
self.l1 = tf.nn.relu(tf.matmul(self.s, self.w_l1) + self.b_l1)
self.w_pi1 = tf.Variable(tf.random_normal([n_units,env.n_actions],stddev=0.1))
self.b_pi1 = tf.Variable(tf.random_normal([env.n_actions],stddev=0.1))
self.actions_prob = tf.nn.softmax(tf.matmul(self.l1, self.w_pi1) + self.b_pi1)
with tf.variable_scope('exp_v'):
log_prob = tf.log(self.actions_prob[0, self.a])
self.exp_v = tf.reduce_mean(log_prob * self.td_error)
with tf.variable_scope('trainActor'):
self.train_op = tf.train.AdamOptimizer(learning_rate).minimize(-self.exp_v)
def learn(self, sess, s, a, td):
s = s[np.newaxis, :]
feed_dict = {self.s: s, self.a: a, self.td_error: td}
_, exp_v = sess.run([self.train_op, self.exp_v], feed_dict)
return exp_v
def calc_action_probs(self, sess, s):
s = s[np.newaxis, :]
probs = sess.run(self.actions_prob, {self.s: s}) # get probabilities for all actions
return probs
class Critic(object):
def __init__(self, env, n_units, learning_rate, gamma, agent_idx,
critic_variant = Critic_Variant.INDEPENDENT):
self.critic_variant = critic_variant
self.env = env
self.s = tf.placeholder(tf.float32, [1, env.n_features], "state")
self.v_ = tf.placeholder(tf.float32, [1, 1], "v_next")
self.r = tf.placeholder(tf.float32, None, 'r')
if self.critic_variant is Critic_Variant.CENTRALIZED:
self.act_probs = tf.placeholder(tf.float32, shape=[1, env.n_actions * env.n_players], name="act_probs")
self.nn_inputs = tf.concat([self.s,self.act_probs],axis=1)
else:
self.nn_inputs = self.s
with tf.variable_scope('Critic'):
l1 = tf.layers.dense(
inputs=self.nn_inputs,
units=n_units, # number of hidden units
activation=tf.nn.relu, # None
kernel_initializer=tf.random_normal_initializer(0., .1), # weights
bias_initializer=tf.constant_initializer(0.1), # biases
name='l1'+str(agent_idx)
)
self.v = tf.layers.dense(
inputs=l1,
units=1, # output units
activation=None,
kernel_initializer=tf.random_normal_initializer(0., .1), # weights
bias_initializer=tf.constant_initializer(0.1), # biases
name='V'+str(agent_idx)
)
with tf.variable_scope('squared_TD_error'):
self.td_error = self.r + gamma * self.v_ - self.v
self.loss = tf.square(self.td_error)
with tf.variable_scope('trainCritic'):
self.train_op = tf.train.AdamOptimizer(learning_rate).minimize(self.loss)
def pass_agent_list(self, agent_list):
self.agent_list = agent_list
def learn(self, sess, s, r, s_, *args):
s,s_ = s.astype(np.float32), s_.astype(np.float32)
if self.critic_variant is Critic_Variant.CENTRALIZED:
if args:
obslist = args[0]
obs_list = args[1]
act_probs = np.hstack([agent.calc_action_probs(obslist[idx]) for idx, agent in enumerate(self.agent_list)])
act_probs_ = np.hstack([agent.calc_action_probs(obs_list[idx]) for idx, agent in enumerate(self.agent_list)])
else:
act_probs = np.hstack([agent.calc_action_probs(s) for idx, agent in enumerate(self.agent_list)])
act_probs_ = np.hstack([agent.calc_action_probs(s_) for idx, agent in enumerate(self.agent_list)])
nn_inputs = np.hstack([s[np.newaxis, :], act_probs])
nn_inputs_ = np.hstack([s_[np.newaxis, :], act_probs_])
else:
nn_inputs, nn_inputs_ = s[np.newaxis, :], s_[np.newaxis, :]
v_ = sess.run(self.v, {self.nn_inputs: nn_inputs_})
td_error, _ = sess.run([self.td_error, self.train_op],
{self.nn_inputs: nn_inputs, self.v_: v_, self.r: r})
return td_error
class Simple_Agent(Agent): #plays games with 2 actions, using a single parameter
def __init__(self, env, learning_rate=0.001, n_units_critic = 20, gamma = 0.95, agent_idx = 0, critic_variant = Critic_Variant.INDEPENDENT):
super().__init__(env, learning_rate, gamma, agent_idx)
self.s = tf.placeholder(tf.float32, [1, env.n_features], "state") # dummy variable
self.a = tf.placeholder(tf.int32, None, "act")
self.td_error = tf.placeholder(tf.float32, None, "td_error") # TD_error
with tf.variable_scope('Actor'):
self.theta = tf.Variable(tf.random_normal([1],mean = -2, stddev=0.5))
self.actions_prob = tf.expand_dims(tf.concat([1-tf.sigmoid(self.theta),tf.sigmoid(self.theta)],0),0)
with tf.variable_scope('exp_v'):
self.log_prob = tf.log(self.actions_prob[0,self.a])
self.g_log_pi = tf.gradients(self.log_prob,self.theta)
self.exp_v = tf.reduce_mean(self.log_prob * self.td_error)
with tf.variable_scope('trainActor'):
self.train_op = tf.train.AdamOptimizer(learning_rate).minimize(-self.exp_v)
self.critic = Critic(env, n_units_critic, learning_rate, gamma, agent_idx, critic_variant)
self.sess.run(tf.global_variables_initializer())
def learn(self, s, a, r, s_, done = False, *args):
if done:
pass
else:
td = self.critic.learn(self.sess,s,r,s_, *args)
feed_dict = {self.a: a, self.td_error: td}
_, exp_v = self.sess.run([self.train_op, self.exp_v], feed_dict)
def __str__(self):
return 'Simple_Agent_'+str(self.agent_idx)
def calc_action_probs(self, s):
s = s[np.newaxis, :]
probs = self.sess.run(self.actions_prob)
return probs
def pass_agent_list(self, agent_list):
self.critic.pass_agent_list(agent_list)
def get_state_variable(self):
return self.s
def calc_g_log_pi(self,s,a):
return self.sess.run(self.g_log_pi,feed_dict={self.s:s,self.a:a})