def __init__(self, history_size, num_layers, units_per_layer, lr, obs_n_shape, act_shape, act_type,
gumbel_temperature, q_network, agent_index, noise, use_ounoise, temporal_mode):
"""
Implementation of the policy network, with optional gumbel softmax activation at the final layer.
"""
self.num_layers = num_layers
self.lr = lr
self.history_size = history_size
self.obs_n_shape = obs_n_shape
self.act_shape = act_shape
self.act_type = act_type
if act_type is Discrete:
self.use_gumbel = True
else:
self.use_gumbel = False
self.use_ounoise = use_ounoise
self.gumbel_temperature = gumbel_temperature
self.q_network = q_network
self.agent_index = agent_index
self.clip_norm = 0.5
self.noise = noise
self.noise_mode = OUNoise(act_shape[0], scale=1.0)
self.temporal_mode = temporal_mode
self.optimizer = tf.keras.optimizers.Adam(lr=self.lr)
### set up network structure
self.obs_input = tf.keras.layers.Input(shape=(self.history_size, self.obs_n_shape[agent_index][0]))
self.temporal_state = None
if self.temporal_mode.lower() == "rnn":
self.temporal_state = tf.keras.layers.GRU(units_per_layer)
elif self.temporal_mode.lower() == "attention":
self.temporal_state = SelfAttention(activation=tf.keras.layers.LeakyReLU(alpha=0.1))
else:
raise RuntimeError(
"Temporal Information Layer should be rnn or attention but %s found!" % self.temporal_mode)
self.hidden_layers = []
for idx in range(num_layers):
layer = tf.keras.layers.Dense(units_per_layer, activation='relu',
name='ag{}pol_hid{}'.format(agent_index, idx))
self.hidden_layers.append(layer)
if self.use_gumbel:
self.output_layer = tf.keras.layers.Dense(self.act_shape, activation='linear',
name='ag{}pol_out{}'.format(agent_index, idx))
else:
self.output_layer = tf.keras.layers.Dense(self.act_shape, activation='tanh',
name='ag{}pol_out{}'.format(agent_index, idx))
# connect layers
x = self.obs_input
x = self.temporal_state(x)
if self.temporal_mode.lower() == "attention":
x = tf.keras.layers.Lambda(lambda x: x[:, -1])(x)
for layer in self.hidden_layers:
x = layer(x)
x = self.output_layer(x)
self.model = tf.keras.Model(inputs=[self.obs_input], outputs=[x])
def __init__(self, num_layers, units_per_layer, lr, obs_n_shape, act_shape,
act_type, gumbel_temperature, q_network, agent_index, noise,
use_ounoise):
"""
Implementation of the policy network, with optional gumbel softmax activation at the final layer.
"""
self.num_layers = num_layers
self.lr = lr
self.obs_n_shape = obs_n_shape
self.act_shape = act_shape
self.act_type = act_type
if act_type is Discrete:
self.use_gumbel = True
else:
self.use_gumbel = False
self.use_ounoise = use_ounoise
self.gumbel_temperature = gumbel_temperature
self.q_network = q_network
self.agent_index = agent_index
self.clip_norm = 0.5
self.noise = noise
self.noise_mode = OUNoise(act_shape[0], scale=1.0)
self.optimizer = tf.keras.optimizers.Adam(lr=self.lr)
### set up network structure
self.obs_input = tf.keras.layers.Input(
shape=self.obs_n_shape[agent_index])
self.hidden_layers = []
for idx in range(num_layers):
layer = tf.keras.layers.Dense(units_per_layer,
activation='relu',
name='ag{}pol_hid{}'.format(
agent_index, idx))
self.hidden_layers.append(layer)
if self.use_gumbel:
self.output_layer = tf.keras.layers.Dense(
self.act_shape,
activation='linear',
name='ag{}pol_out{}'.format(agent_index, idx))
else:
self.output_layer = tf.keras.layers.Dense(
self.act_shape,
activation='tanh',
name='ag{}pol_out{}'.format(agent_index, idx))
# connect layers
x = self.obs_input
for layer in self.hidden_layers:
x = layer(x)
x = self.output_layer(x)
self.model = tf.keras.Model(inputs=[self.obs_input], outputs=[x])
def main(arglist):
global no_actions, no_features, no_agents
env = u.make_env(arglist.scenario, arglist.no_agents)
obs_shape_n = env.observation_space
act_shape_n = env.action_space
act_shape_n = u.space_n_to_shape_n(act_shape_n)
no_agents = env.n
batch_size = arglist.batch_size
no_neighbors = arglist.no_neighbors
k_lst = list(range(no_neighbors + 2))[2:] # [2,3]
u.create_seed(arglist.seed)
noise_mode = OUNoise(act_shape_n[0], scale=1.0)
noise = 0.1
reduction_noise = 0.999
# Velocity.x Velocity.y Pos.x Pos.y {Land.Pos.x Land.Pos.y}*10 {Ent.Pos.x Ent.Pos.y}*9
no_features = obs_shape_n[0].shape[0]
no_actions = act_shape_n[0][0]
model, model_t = __build_conf()
optimizer = AdamW(learning_rate=arglist.lr, weight_decay=1e-5)
# Results
episode_rewards = [0.0] # sum of rewards for all agents
result_path = os.path.join("results", arglist.exp_name)
res = os.path.join(result_path, " %s.csv" % arglist.exp_name)
if not os.path.exists(result_path):
os.makedirs(result_path)
replay_buffer = ReplayBuffer(arglist.max_buffer_size) # Init Buffer
episode_step = 0
train_step = 0
t_start = time.time()
obs_n = env.reset()
adj = u.get_adj(obs_n, k_lst, no_agents, is_gcn=True)
print('Starting iterations...')
while True:
episode_step += 1
terminal = (episode_step >= arglist.max_episode_len)
if episode_step % 3 == 0:
adj = u.get_adj(obs_n, k_lst, no_agents, is_gcn=True)
predictions = get_predictions(u.to_tensor(np.array(obs_n)), adj, model)
actions = get_actions(predictions, noise, noise_mode)
# Observe next state, reward and done value
new_obs_n, rew_n, done_n, _ = env.step(actions)
done = all(done_n) or terminal
cooperative_reward = rew_n[0]
# Store the data in the replay memory
replay_buffer.add(obs_n, adj, actions, cooperative_reward, new_obs_n,
done)
obs_n = new_obs_n
episode_rewards[-1] += cooperative_reward
if done or terminal:
obs_n = env.reset()
episode_step = 0
episode_rewards.append(0)
# increment global step counter
train_step += 1
# for displaying learned policies
if arglist.display:
time.sleep(0.1)
env.render()
continue
# Train the models
train_cond = not arglist.display
if train_cond and len(replay_buffer) > arglist.batch_size:
if len(
episode_rewards
) % arglist.update_rate == 0: # only update every 30 episodes
for _ in range(arglist.update_times):
state, adj_n, actions, rewards, new_state, dones = replay_buffer.sample(
batch_size)
noise *= reduction_noise
# Calculate TD-target
with tf.GradientTape() as tape:
target_q_values = model_t([new_state, adj_n])
# Apply max(Q) to obtain the TD-target
target_q_tot = tf.reduce_max(target_q_values, axis=-1)
# Apply VDN to reduce the agent-dimension
max_q_tot = tf.reduce_sum(target_q_tot, axis=-1)
y = rewards + (1. - dones) * arglist.gamma * max_q_tot
# Predictions
action_one_hot = tf.one_hot(
tf.argmax(actions, axis=2, name='action_one_hot'),
no_actions)
q_values = model([state, adj_n])
q_tot = tf.reduce_sum(q_values * action_one_hot,
axis=-1,
name='q_acted')
pred = tf.reduce_sum(q_tot, axis=1)
if "huber" in arglist.loss_type:
loss = tf.reduce_sum(
u.huber_loss(pred, tf.stop_gradient(y)))
elif "mse" in arglist.loss_type:
loss = tf.losses.mean_squared_error(
pred, tf.stop_gradient(y))
else:
raise RuntimeError(
"Loss function should be either Huber or MSE. %s found!"
% arglist.loss_type)
gradients = tape.gradient(loss,
model.trainable_variables)
local_clipped = u.clip_by_local_norm(gradients, 0.1)
optimizer.apply_gradients(
zip(local_clipped, model.trainable_variables))
tf.saved_model.save(model, result_path)
# display training output
if train_step % arglist.save_rate == 0:
# eval_reward = get_eval_reward(env, model)
with open(res, "a+") as f:
mes_dict = {
"steps":
train_step,
"episodes":
len(episode_rewards),
"train_episode_reward":
np.round(np.mean(episode_rewards[-arglist.save_rate:]),
3),
# "eval_episode_reward": np.round(np.mean(eval_reward), 3),
"time":
round(time.time() - t_start, 3)
}
print(mes_dict)
for item in list(mes_dict.values()):
f.write("%s\t" % item)
f.write("\n")
f.close()
t_start = time.time()
# train target model
if arglist.soft_update:
weights = model.get_weights()
target_weights = model_t.get_weights()
for w in range(len(weights)):
target_weights[w] = arglist.tau * weights[w] + (
1 - arglist.tau) * target_weights[w]
model_t.set_weights(target_weights)
elif terminal and train_step % 200 == 0:
model_t.set_weights(model.get_weights())