def get_eval_reward(env, model):
k_lst = list(range(arglist.no_neighbors + 2))[2:] # [2,3]
reward_total = []
for _ in range(3):
obs_n = env.reset()
adj = u.get_adj(obs_n, k_lst, no_agents, is_gcn=True)
reward = 0
for i in range(arglist.max_episode_len):
predictions = get_predictions(u.to_tensor(np.array(obs_n)), adj,
model)
predictions = tf.squeeze(predictions, axis=0)
# actions = [tf.argmax(prediction, axis=-1).numpy() for prediction in predictions]
# Observe next state, reward and done value
new_obs_n, rew_n, done_n, _ = env.step(predictions.numpy())
adj = u.get_adj(new_obs_n, k_lst, no_agents, is_gcn=True)
obs_n = new_obs_n
reward += rew_n[0]
reward_total.append(reward)
return reward_total
def main(arglist):
# Global variables
global num_actions, feature_dim, no_agents
# Create environment
env = u.make_env(arglist.scenario, no_agents=arglist.no_agents)
env.discrete_action_input = True
obs_shape_n = env.observation_space
no_agents = env.n
no_neighbors = arglist.num_neighbors
u.create_seed(3)
k_lst = list(range(no_neighbors + 2))[2:] # [2,3]
# Velocity.x Velocity.y Pos.x Pos.y {Land.Pos.x Land.Pos.y}*10 {Ent.Pos.x Ent.Pos.y}*9
num_features = obs_shape_n[0].shape[0]
num_actions = env.action_space[0].n
feature_dim = num_features # the size of node features
result_path = "results/mixing/" + arglist.exp_name
if not os.path.exists(result_path):
os.makedirs(result_path)
model = keras.models.load_model(result_path)
res = "asymptotic/" + arglist.exp_name
if not os.path.exists(res):
os.makedirs(res)
res = "asymptotic/" + arglist.exp_name + "/testing_rewards.csv"
episode_rewards = []
while True:
obs_n = env.reset()
if arglist.use_gat or arglist.use_gcn:
adj = u.get_adj(obs_n,
k_lst,
no_agents,
is_gat=arglist.use_gat,
is_gcn=arglist.use_gcn)
else:
adj = None
if arglist.use_rnn:
obs_n = u.reshape_state(obs_n, arglist.history_size)
episode_rewards.append(0)
for i in range(arglist.max_episode_len):
predictions = get_predictions(u.to_tensor(np.array(obs_n)), adj,
model)
predictions = tf.squeeze(predictions, axis=0)
# print("predictions: %s" % tf.shape(predictions))
actions = [
tf.argmax(prediction, axis=-1).numpy()
for prediction in predictions
]
# Observe next state, reward and done value
new_obs_n, rew_n, done_n, _ = env.step(actions)
if arglist.use_gat or arglist.use_gcn:
adj = u.get_adj(obs_n,
k_lst,
no_agents,
is_gat=arglist.use_gat,
is_gcn=arglist.use_gcn)
if arglist.use_rnn:
new_obs_n = u.refresh_history(np.copy(obs_n), new_obs_n)
obs_n = new_obs_n
episode_rewards[-1] += sum(rew_n)
# for displaying learned policies
if arglist.display:
time.sleep(0.5)
# print("Reward is %.3f" % sum(rew_n))
env.render()
continue
with open(res, "a+") as f:
mes_dict = {
"episodes": len(episode_rewards),
"train_episode_reward": np.round(np.mean(episode_rewards[-1]),
3)
}
print(mes_dict)
for item in list(mes_dict.values()):
f.write("%s\t" % item)
f.write("\n")
f.close()
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())
def main(display, restore_fp, arglist):
no_agents = arglist.no_agents
u.create_seed(arglist.seed)
env = make_env(arglist.scenario)
if restore_fp is not None:
with open(os.path.join(restore_fp, 'args.pkl'), 'rb') as f:
arglist = pickle.load(f)
arglist.display = False
arglist.use_ounoise = False
arglist.save_rate = 10
arglist.restore_fp = True
temp = os.path.join(*(restore_fp.split(os.path.sep)[1:]))
arglist.exp_name = os.path.join('evaluation', temp)
# TODO: no_gnn_neurons
arglist.no_gnn_neurons = arglist.no_critic_neurons
restore_fp = os.path.join(restore_fp, 'models')
logger = RLLogger(env.n, env.n_adversaries, arglist.save_rate, arglist)
obs_shape_n = u.space_n_to_shape_n(env.observation_space)
act_shape_n = u.space_n_to_shape_n(env.action_space)
model_path = os.path.join("results", arglist.exp_name, 'models')
os.makedirs(model_path, exist_ok=True)
k_lst = list(range(arglist.no_neighbors + 2))[2:]
critic = MADDPGCriticNetwork(arglist.no_layers,
arglist.no_critic_neurons,
arglist.lr,
obs_shape_n,
act_shape_n,
wd=1e-5)
critic_target = MADDPGCriticNetwork(arglist.no_layers,
arglist.no_critic_neurons,
arglist.lr,
obs_shape_n,
act_shape_n,
wd=1e-5)
critic_target.model.set_weights(critic.model.get_weights())
agents = get_agents(env, arglist.lr, arglist.no_layers,
arglist.no_actor_neurons, arglist.tau, arglist.noise,
arglist.use_ounoise, logger)
obs_n = env.reset()
replay_buffer = EfficientReplayBuffer(int(arglist.max_buffer_size),
no_agents, obs_shape_n,
act_shape_n) # Init Buffer
# Load previous results if necessary
if restore_fp is not None:
print('Loading previous state...')
for ag_idx, agent in enumerate(agents):
fp = os.path.join(restore_fp, 'agent_{}'.format(ag_idx))
agent.load(fp)
critic.load(restore_fp + '/critic.h5')
critic_target.load(restore_fp + '/critic_target.h5')
print('Starting iterations...')
while True:
logger.episode_step += 1
action_n = [
agent.action(obs.astype(np.float32)).numpy()
for agent, obs in zip(agents, obs_n)
]
new_obs_n, rew_n, done_n, _ = env.step(action_n)
cooperative_reward = rew_n[0]
terminal = (logger.episode_step >= arglist.max_episode_len)
done = all(done_n) or terminal
# collect experience
replay_buffer.add(obs_n, action_n, cooperative_reward, new_obs_n, done)
obs_n = new_obs_n
if done:
obs_n = env.reset()
episode_step = 0
logger.record_episode_end(agents, arglist.display)
for ag_idx, rew in enumerate(rew_n):
logger.cur_episode_reward += cooperative_reward
logger.agent_rewards[ag_idx][-1] += cooperative_reward
logger.train_step += 1
# for displaying learned policies
if arglist.display:
time.sleep(0.1)
env.render()
train_cond = not display
if restore_fp is None:
if train_cond and len(replay_buffer) > arglist.batch_size:
if logger.episode_count % arglist.update_rate == 0: # only update every 30 episodes
for _ in range(arglist.update_times):
# Sample: Shapes --> (no-agents, batch_size, features)
state, actions, rewards, new_state, dones = replay_buffer.sample(
arglist.batch_size)
target_act_next = [
a.target_action(obs)
for a, obs in zip(agents, new_state)
]
adjacency = [
u.get_adj(obs, k_lst, no_agents, is_gat=True)
for obs in np.swapaxes(state, 1, 0)
]
adjacency = np.array(
adjacency
) # shape: (batch_size, no_agents, no_agents)
target_q_next = critic_target.predict(
new_state, target_act_next, adjacency)
q_train_target = rewards + (
1. - dones) * arglist.gamma * target_q_next
loss, td_loss = critic.train_step(
state, actions, adjacency, q_train_target)
logger.save_logger("critic_loss", np.mean(td_loss),
logger.train_step, 0)
update_target_networks(critic, critic_target)
critic.save(model_path + '/critic.h5')
critic_target.save(model_path + '/critic_target.h5')
for agent in agents:
pol_loss = agent.update(state, actions, adjacency,
critic, logger.train_step)
# saves logger outputs to a file similar to the way in the original MADDPG implementation
if len(logger.episode_rewards) > arglist.no_episodes:
logger.experiment_end()
return logger.get_sacred_results()