/
ProcessAgent.py
133 lines (111 loc) · 5.18 KB
/
ProcessAgent.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
# Copyright (c) 2016, NVIDIA CORPORATION. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
from datetime import datetime
from multiprocessing import Process, Queue, Value
import numpy as np
import time
from Config import Config
from Environment import Environment
from Experience import Experience
class ProcessAgent(Process):
def __init__(self, id, prediction_q, training_q, episode_log_q):
super(ProcessAgent, self).__init__()
self.id = id
self.prediction_q = prediction_q
self.training_q = training_q
self.episode_log_q = episode_log_q
self.env = Environment()
self.num_actions = self.env.get_num_actions()
self.actions = np.arange(self.num_actions)
self.discount_factor = Config.DISCOUNT
# one frame at a time
self.wait_q = Queue(maxsize=1)
self.exit_flag = Value('i', 0)
@staticmethod
def _accumulate_rewards(experiences, discount_factor, terminal_reward):
reward_sum = terminal_reward
for t in reversed(range(0, len(experiences)-1)):
r = np.clip(experiences[t].reward, Config.REWARD_MIN, Config.REWARD_MAX)
reward_sum = discount_factor * reward_sum + r
experiences[t].reward = reward_sum
return experiences[:-1]
def convert_data(self, experiences):
x_ = np.array([exp.state for exp in experiences])
a_ = np.eye(self.num_actions)[np.array([exp.action for exp in experiences])].astype(np.float32)
r_ = np.array([exp.reward for exp in experiences])
return x_, r_, a_
def predict(self, state):
# put the state in the prediction q
self.prediction_q.put((self.id, state))
# wait for the prediction to come back
p, v = self.wait_q.get()
return p, v
def select_action(self, prediction):
if Config.PLAY_MODE:
action = np.argmax(prediction)
else:
action = np.random.choice(self.actions, p=prediction)
return action
def run_episode(self):
self.env.reset()
done = False
experiences = []
time_count = 0
reward_sum = 0.0
while not done:
# very first few frames
if self.env.current_state is None:
self.env.step(0) # 0 == NOOP
continue
prediction, value = self.predict(self.env.current_state)
action = self.select_action(prediction)
reward, done = self.env.step(action)
reward_sum += reward
exp = Experience(self.env.previous_state, action, prediction, reward, done)
experiences.append(exp)
if done or time_count == Config.TIME_MAX:
terminal_reward = 0 if done else value
updated_exps = ProcessAgent._accumulate_rewards(experiences, self.discount_factor, terminal_reward)
x_, r_, a_ = self.convert_data(updated_exps)
yield x_, r_, a_, reward_sum
# reset the tmax count
time_count = 0
# keep the last experience for the next batch
experiences = [experiences[-1]]
reward_sum = 0.0
time_count += 1
def run(self):
# randomly sleep up to 1 second. helps agents boot smoothly.
time.sleep(np.random.rand())
np.random.seed(np.int32(time.time() % 1 * 1000 + self.id * 10))
while self.exit_flag.value == 0:
total_reward = 0
total_length = 0
for x_, r_, a_, reward_sum in self.run_episode():
total_reward += reward_sum
total_length += len(r_) + 1 # +1 for last frame that we drop
self.training_q.put((x_, r_, a_))
self.episode_log_q.put((datetime.now(), total_reward, total_length))