def __init__(self,
name,
before_com_model,
channel,
after_com_model,
critic_mlp_model,
obs_shape_n,
act_space_n,
args,
local_q_func=False):
self.name = name
self.n = len(obs_shape_n)
self.args = args
self.alpha = 0.5
obs_ph_n = []
for i in range(self.n):
obs_ph_n.append(
U.BatchInput(obs_shape_n[i],
name="observation_" + str(i)).get())
# Create all the functions necessary to train the model
self.q_train, self.q_update, self.q_debug = q_train(
scope=self.name,
make_obs_ph_n=obs_ph_n,
act_space_n=act_space_n,
q_func=critic_mlp_model,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr),
grad_norm_clipping=0.5,
local_q_func=local_q_func,
num_units=args.num_units,
)
self.act, self.p_train, self.p_update, self.p_debug = self.p_train_function(
scope=self.name,
make_obs_ph_n=obs_ph_n,
act_space_n=act_space_n,
before_com_func=before_com_model,
channel=channel,
after_com_func=after_com_model,
q_func=critic_mlp_model,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr),
grad_norm_clipping=0.5,
local_q_func=local_q_func,
num_units=args.num_units,
beta=args.beta,
ibmac_com=args.ibmac_com,
)
# Create experience buffer
self.replay_buffer = ReplayBuffer(1e6)
# self.max_replay_buffer_len = 50 * args.max_episode_len
self.max_replay_buffer_len = args.batch_size * args.max_episode_len
self.replay_sample_index = None
self.message_1_for_record = []
def __init__(self,
name,
model,
obs_shape_n,
act_space_n,
agent_index,
args,
local_q_func=False):
self.name = name
self.n = len(obs_shape_n)
self.agent_index = agent_index
self.args = args
obs_ph_n = []
for i in range(self.n):
obs_ph_n.append(
U.BatchInput(obs_shape_n[i],
name="observation" + str(i)).get())
# Create all the functions necessary to train the model
self.q_train, self.q_update, self.q_debug = q_train(
scope=self.name,
make_obs_ph_n=obs_ph_n,
act_space_n=act_space_n,
q_index=agent_index,
q_func=model,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr),
grad_norm_clipping=0.5,
local_q_func=local_q_func,
num_units=args.num_units)
self.act, self.p_train, self.p_update, self.p_debug = p_train(
scope=self.name,
make_obs_ph_n=obs_ph_n,
act_space_n=act_space_n,
p_index=agent_index,
p_func=model,
q_func=model,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr),
grad_norm_clipping=0.5,
local_q_func=local_q_func,
num_units=args.num_units)
# Create experience buffer
self.replay_buffer = ReplayBuffer(1e6)
self.max_replay_buffer_len = args.batch_size * args.max_episode_len
self.replay_sample_index = None