"""
Multi-Agent Deep Deterministic Policy Gradient
Paper link:
https://proceedings.neurips.cc/paper/2017/file/68a9750337a418a86fe06c1991a1d64c-Paper.pdf
Implementation: TensorFlow 2.X
Trick: Parameter sharing for all agents, with agents' one-hot IDs as actor-critic's inputs.
"""
from argparse import Namespace
from operator import itemgetter
from xuance.common import List
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners import LearnerMAS
[docs]
class MADDPG_Learner(LearnerMAS):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
super(MADDPG_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
self.build_optimizer()
self.gamma = config.gamma
self.tau = config.tau
[docs]
def build_optimizer(self):
if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips.
self.optimizer = {
key: {'actor': tk.optimizers.legacy.Adam(self.config.learning_rate_actor),
'critic': tk.optimizers.legacy.Adam(self.config.learning_rate_critic)}
for key in self.model_keys}
else:
self.optimizer = {
key: {'actor': tk.optimizers.Adam(self.config.learning_rate_actor),
'critic': tk.optimizers.Adam(self.config.learning_rate_critic)}
for key in self.model_keys}
@tf.function
def forward_fn(self, batch_size, bs, obs, obs_joint, actions, actions_joint, rewards,
obs_next, next_obs_joint, terminals, IDs, agent_mask):
info_train = {}
gradients_a, gradients_c = {}, {}
with tf.GradientTape(persistent=True) as tape:
# update critic
_, actions_next = self.policy.Atarget(next_observation=obs_next, agent_ids=IDs)
if self.use_parameter_sharing:
key = self.model_keys[0]
actions_next_joint = tf.reshape(tf.reshape(actions_next[key], [batch_size, self.n_agents, -1]),
[batch_size, -1])
else:
actions_next_joint = tf.reshape(tf.concat(itemgetter(*self.model_keys)(actions_next), -1),
[batch_size, -1])
_, q_eval = self.policy.Qpolicy(joint_observation=obs_joint, joint_actions=actions_joint, agent_ids=IDs)
_, q_next = self.policy.Qtarget(joint_observation=next_obs_joint, joint_actions=actions_next_joint,
agent_ids=IDs)
for key in self.model_keys:
mask_values = agent_mask[key]
q_eval_a = tf.reshape(q_eval[key], [bs])
q_next_i = tf.reshape(q_next[key], [bs])
q_target = rewards[key] + (1 - terminals[key]) * self.gamma * q_next_i
td_error = (q_eval_a - tf.stop_gradient(q_target)) * mask_values
loss_c = tf.reduce_sum(td_error ** 2) / tf.reduce_sum(mask_values)
gradients_c[key] = tape.gradient(loss_c, self.policy.critic_trainable_variables(key))
if self.use_grad_clip:
gradients_c[key], _ = tf.clip_by_global_norm(gradients_c[key], clip_norm=self.grad_clip_norm)
self.optimizer[key]['critic'].apply_gradients(zip(gradients_c[key],
self.policy.critic_trainable_variables(key)))
else:
self.optimizer[key]['critic'].apply_gradients(zip(gradients_c[key],
self.policy.critic_trainable_variables(key)))
info_train.update({f"{key}/loss_critic": loss_c,
f"{key}/predictQ": tf.math.reduce_mean(q_eval[key])})
# Update actor
_, actions_eval = self.policy(observation=obs, agent_ids=IDs)
for key in self.model_keys:
mask_values = agent_mask[key]
if self.use_parameter_sharing:
act_eval = tf.reshape(tf.reshape(actions_eval[key], [batch_size, self.n_agents, -1]),
[batch_size, -1])
else:
a_joint = {k: actions_eval[k] if k == key else actions[k] for k in self.agent_keys}
act_eval = tf.reshape(tf.concat(itemgetter(*self.agent_keys)(a_joint), axis=-1),
[batch_size, -1])
_, q_policy = self.policy.Qpolicy(joint_observation=obs_joint, joint_actions=act_eval,
agent_ids=IDs, agent_key=key)
q_policy_i = tf.reshape(q_policy[key], [bs])
loss_a = -tf.reduce_sum(q_policy_i * mask_values) / tf.reduce_sum(mask_values)
gradients_a[key] = tape.gradient(loss_a, self.policy.actor_trainable_variables(key))
if self.use_grad_clip:
gradients_a[key], _ = tf.clip_by_global_norm(gradients_a[key], clip_norm=self.grad_clip_norm)
self.optimizer[key]['actor'].apply_gradients(zip(gradients_a[key],
self.policy.actor_trainable_variables(key)))
else:
self.optimizer[key]['actor'].apply_gradients(zip(gradients_a[key],
self.policy.actor_trainable_variables(key)))
info_train.update({f"{key}/loss_actor": loss_a})
self.policy.soft_update(self.tau)
return info_train
@tf.function
def learn(self, *inputs):
if self.distributed_training:
info_train = self.policy.mirrored_strategy.run(self.forward_fn, args=inputs)
return info_train[0]
else:
return self.forward_fn(*inputs)
[docs]
def update(self, sample):
self.iterations += 1
# prepare training data
sample_Tensor = self.build_training_data(sample,
use_parameter_sharing=self.use_parameter_sharing,
use_actions_mask=False)
batch_size = sample_Tensor['batch_size']
obs = sample_Tensor['obs']
actions = sample_Tensor['actions']
obs_next = sample_Tensor['obs_next']
rewards = sample_Tensor['rewards']
terminals = sample_Tensor['terminals']
agent_mask = sample_Tensor['agent_mask']
IDs = sample_Tensor['agent_ids']
if self.use_parameter_sharing:
key = self.model_keys[0]
bs = batch_size * self.n_agents
obs_joint = tf.reshape(obs[key], [batch_size, -1])
next_obs_joint = tf.reshape(obs_next[key], [batch_size, -1])
actions_joint = tf.reshape(actions[key], [batch_size, -1])
rewards[key] = tf.reshape(rewards[key], [batch_size * self.n_agents])
terminals[key] = tf.reshape(terminals[key], [batch_size * self.n_agents])
else:
bs = batch_size
obs_joint = tf.reshape(tf.concat(itemgetter(*self.agent_keys)(obs), axis=-1), [batch_size, -1])
next_obs_joint = tf.reshape(tf.concat(itemgetter(*self.agent_keys)(obs_next), axis=-1), [batch_size, -1])
actions_joint = tf.reshape(tf.concat(itemgetter(*self.agent_keys)(actions), axis=-1), [batch_size, -1])
info = self.callback.on_update_start(self.iterations, method="update",
policy=self.policy, sample_Tensor=sample_Tensor, bs=bs,
obs_joint=obs_joint, next_obs_joint=next_obs_joint,
actions_joint=actions_joint)
info_train = self.learn(batch_size, bs, obs, obs_joint, actions, actions_joint, rewards,
obs_next, next_obs_joint, terminals, IDs, agent_mask)
for k, v in info_train.items():
info_train[k] = v.numpy()
info.update(info_train)
info.update(self.callback.on_update_end(self.iterations, method="update_rnn", policy=self.policy, info=info))
return info