-
Notifications
You must be signed in to change notification settings - Fork 1
/
RAC.py
171 lines (140 loc) · 6.13 KB
/
RAC.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
from __future__ import print_function
import argparse
import configparser
import csv
import logging
import os
import sys
from datetime import datetime
import gym
import gym.wrappers as wrappers
import numpy as np
from Agent.RACAgent import RACAgent
class RAC(object):
def __init__(self):
self.__initialise_agents()
def __initialise_agents(self):
self.agent = RACAgent(props.getint('feature', 'dimension'),
props.getint('policy', 'num_actions'),
props.getint('state', 'state_length'),
props.getint('state', 'feature_size'))
def execute_algorithm(self):
max_episodes = props.getint('train', 'max_episodes')
max_steps = props.getint('train', 'max_steps')
step_size = props.getint('train', 'step_size')
update_count = 0
for i in range(max_episodes):
t_start = datetime.now()
steps = 0
state = env.reset()
terminal_reached = False
while not terminal_reached and steps < max_steps:
# Predict action
state_feature = self.agent.feature.phi(state)
action, distribution = self.agent.policy.get_action(state_feature)
# take action and observe reward
next_state, reward, done, info = env.step(action)
for x in range(step_size - 1):
if done:
terminal_reached = True
break
next_state, reward2, done, info = env.step(action)
reward += reward2
if done:
terminal_reached = True
steps += 1
self.agent.update_parameters(state, action, reward, next_state)
state = next_state
if i % 100 == 0:
# test agent
test_agent(self.agent, i)
logger.debug("Completed Iteration %d. Time taken: %f", i, (datetime.now() - t_start).total_seconds())
def test_agent(agent, episode_count):
env_test = gym.make(args.env_id)
if display_game:
outdir = 'videos/tmp/neat-data/{0}-{1}'.format(env_test.spec.id, str(datetime.now()))
env_test = wrappers.Monitor(env_test, directory=outdir, force=True)
logger.debug("Generating best agent result: %d", episode_count)
t_start = datetime.now()
test_episodes = props.getint('test', 'test_episodes')
step_size = props.getint('test', 'step_size')
avg_steps = []
avg_rewards = []
for i in range(test_episodes):
state = env_test.reset()
terminal_reached = False
steps = 0
rewards = 0
while not terminal_reached:
if display_game:
env.render()
# Predict action
state_feature = agent.feature.phi(state)
action, distribution = agent.policy.get_action(state_feature)
# take action and observe reward
next_state, reward, done, info = env_test.step(action)
for x in range(step_size - 1):
if done:
terminal_reached = True
break
next_state, reward2, done, info = env_test.step(action)
reward += reward2
steps += 1
rewards += reward
state = next_state
if done:
terminal_reached = True
avg_steps.append(steps)
avg_rewards.append(rewards)
average_steps_per_episode = np.sum(avg_steps) / len(avg_steps)
average_rewards_per_episode = np.sum(avg_rewards) / len(avg_rewards)
# save this to file along with the generation number
entry = [episode_count, average_steps_per_episode, average_rewards_per_episode]
with open(r'agent_evaluation-{0}.csv'.format(time), 'a') as file:
writer = csv.writer(file)
writer.writerow(entry)
logger.debug("Finished: evaluating best agent. Time taken: %f", (datetime.now() - t_start).total_seconds())
env_test.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=None)
parser.add_argument('env_id', nargs='?', default='CartPole-v0', help='Select the environment to run')
parser.add_argument('display', nargs='?', default='false', help='Show display of game. true or false')
args = parser.parse_args()
# Call `undo_logger_setup` if you want to undo Gym's logger setup
# and configure things manually. (The default should be fine most
# of the time.)
gym.undo_logger_setup()
time = datetime.now().strftime("%Y%m%d-%H:%M:%S")
logging.basicConfig(filename='log/debug-{0}.log'.format(time),
level=logging.DEBUG, format='[%(asctime)s] %(message)s')
logger = logging.getLogger()
formatter = logging.Formatter('[%(asctime)s] %(message)s')
handler = logging.StreamHandler(sys.stderr)
handler.setFormatter(formatter)
logger.addHandler(handler)
# You can set the level to logging.DEBUG or logging.WARN if you
# want to change the amount of output.
logger.setLevel(logging.DEBUG)
env = gym.make(args.env_id)
logger.debug("action space: %s", env.action_space)
logger.debug("observation space: %s", env.observation_space)
# Load the properties file
local_dir = os.path.dirname(__file__)
logger.debug("Loading Properties")
props = configparser.ConfigParser()
prop_path = os.path.join(local_dir, 'properties/{0}/neatem_properties.ini'.format(env.spec.id))
props.read(prop_path)
logger.debug("Finished: Loading Properties")
agent = RAC()
# Run until the winner from a generation is able to solve the environment
# or the user interrupts the process.
display_game = True if args.display == 'true' else False
try:
agent.execute_algorithm()
except KeyboardInterrupt:
logger.debug("User break.")
finally:
env.close()
# Upload to the scoreboard. We could also do this from another
# logger.info("Successfully ran RandomAgent. Now trying to upload results to the scoreboard. If it breaks, you can always just try re-uploading the same results.")
# gym.upload(outdir)