master_network = AC_Network(
state_size,
s_size_central,
number_of_agents,
action_size, (number_of_agents - 1) * comm_size,
(number_of_agents - 1) * comm_size if spread_messages else comm_size,
'global',
None,
critic_action=critic_action,
critic_comm=critic_comm) # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(
Worker(make_env("simple_spread6", benchmark=True),
i,
state_size,
s_size_central,
action_size,
number_of_agents,
trainer,
model_path,
global_episodes,
display=display and i == 0,
comm=(comm_size != 0),
comm_size_per_agent=comm_size,
spread_messages=spread_messages,
critic_action=critic_action,
critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.
action_size[3],
0,
0,
'global',
None,
'_agentPrey',
critic_action=critic_action,
critic_comm=critic_comm)
master_networks = [predator, predator, predator,
prey] # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(
Worker(make_env("simple_tag"),
i,
state_size,
s_size_central,
action_size,
number_of_agents,
trainer,
model_path,
global_episodes,
display=display and i == 0,
comm=(comm_size != 0),
comm_size_per_agent=comm_size,
spread_messages=spread_messages,
critic_action=critic_action,
critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.
number_of_agents,
action_size[0],
0,
0,
'global',
None,
'_agentPrey',
critic_action=critic_action,
critic_comm=critic_comm)
master_networks = [prey, predator, predator] # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(
Worker(make_env("simple_adversary"),
i,
state_size,
s_size_central,
action_size,
number_of_agents,
trainer,
model_path,
global_episodes,
display=display and i == 0,
comm=(comm_size != 0),
comm_size_per_agent=comm_size,
spread_messages=spread_messages,
critic_action=critic_action,
critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.
os.makedirs(model_path)
with tf.device("/cpu:0"):
global_episodes = tf.Variable(0, dtype=tf.int32, name='global_episodes', trainable=False)
trainer = tf.train.AdamOptimizer(learning_rate=learning_rate)
master_networks = [AC_Network(state_size[i], s_size_central[i], number_of_agents, action_size[i],
(number_of_agents - 1) * comm_size,
(number_of_agents - 1) * comm_size if spread_messages else comm_size,
'global',
None, '_agent'+str(i), critic_action=critic_action, critic_comm=critic_comm)
for i in range(number_of_agents)] # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(Worker(make_env("simple_speaker_listener"), i, state_size, s_size_central,
action_size, number_of_agents, trainer, model_path,
global_episodes,
display=display and i == 0, comm=(comm_size != 0),
comm_size_per_agent=comm_size, spread_messages=spread_messages,
critic_action=critic_action, critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.comm_delivery_failure_chance,
comm_gaussian_noise=FLAGS.comm_gaussian_noise,
comm_jumble_chance=FLAGS.comm_jumble_chance))
saver = tf.train.Saver()
with tf.Session() as sess:
coord = tf.train.Coordinator()
if load_model:
print('Loading Model...')
os.makedirs(model_path)
with tf.device("/cpu:0"):
global_episodes = tf.Variable(0, dtype=tf.int32, name='global_episodes', trainable=False)
trainer = tf.train.AdamOptimizer(learning_rate=learning_rate)
predator = AC_Network(state_size[2], s_size_central[2], number_of_agents, action_size[2],
comm_size, comm_size, 'global',
None, '_agentPredator', critic_action=critic_action, critic_comm=critic_comm)
prey = AC_Network(state_size[1], s_size_central[1], number_of_agents, action_size[1], comm_size, comm_size,
'global', None, '_agentPrey', critic_action=critic_action, critic_comm=critic_comm)
master_networks = [None, prey, predator] # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(Worker(make_env("simple_adversary"), i, state_size, s_size_central,
action_size, number_of_agents, trainer, model_path,
global_episodes,
display=display and i == 0, comm=(comm_size != 0),
comm_size_per_agent=comm_size, spread_messages=spread_messages,
critic_action=critic_action, critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.comm_delivery_failure_chance,
comm_gaussian_noise=FLAGS.comm_gaussian_noise,
comm_jumble_chance=FLAGS.comm_jumble_chance))
saver = tf.train.Saver()
with tf.Session() as sess:
coord = tf.train.Coordinator()
if load_model:
print('Loading Model...')
master_network = AC_Network(
state_size,
s_size_central,
number_of_agents,
action_size, (number_of_agents - 1) * comm_size,
(number_of_agents - 1) * comm_size if spread_messages else comm_size,
'global',
None,
critic_action=critic_action,
critic_comm=critic_comm) # Generate global network
workers = []
# Create worker classes
for i in range(FLAGS.num_slaves):
workers.append(
Worker(make_env("simple_reference"),
i,
state_size,
s_size_central,
action_size,
number_of_agents,
trainer,
model_path,
global_episodes,
display=display and i == 0,
comm=(comm_size != 0),
comm_size_per_agent=comm_size,
spread_messages=spread_messages,
critic_action=critic_action,
critic_comm=critic_comm,
comm_delivery_failure_chance=FLAGS.