"""
COMA: Counterfactual Multi-Agent Policy Gradients
Paper link: https://ojs.aaai.org/index.php/AAAI/article/view/11794
Implementation: TensorFlow 2.X
"""
from tensorflow import one_hot
from argparse import Namespace
from xuance.tensorflow import tf, tk, Module
from xuance.common import List
from xuance.tensorflow.learners.multi_agent_rl.iac_learner import IAC_Learner
[docs]
class COMA_Learner(IAC_Learner):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
config.use_value_clip, config.value_clip_range = False, None
config.use_huber_loss, config.huber_delta = False, None
config.use_value_norm = False
config.vf_coef, config.ent_coef = None, None
super(COMA_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
self.sync_frequency = config.sync_frequency
self.n_actions = {k: self.policy.action_space[k].n for k in self.model_keys}
[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 = {
'actor': tk.optimizers.legacy.Adam(self.config.learning_rate_actor),
'critic': tk.optimizers.legacy.Adam(self.config.learning_rate_critic),
}
else:
self.optimizer = {
'actor': tk.optimizers.Adam(self.config.learning_rate_actor),
'critic': tk.optimizers.Adam(self.config.learning_rate_critic)
}
# @tf.function
[docs]
def forward_fn(self, *args):
bs, batch_size, obs, state, actions, agent_mask, avail_actions, returns, IDs, epsilon = args
with tf.GradientTape(persistent=True) as tape:
_, pi_logits = self.policy(observation=obs, agent_ids=IDs, avail_actions=avail_actions)
if self.use_parameter_sharing:
key = self.model_keys[0]
actions_onehot = {key: one_hot(actions[key], self.n_actions[key])}
else:
IDs = tf.reshape(tf.tile(tf.eye(self.n_agents)[None], [batch_size, 1, 1]), [bs, -1])
actions_onehot = {k: one_hot(actions[k], self.n_actions[k]) for k in self.agent_keys}
_, values_pred = self.policy.get_values(state=state, observation=obs, actions=actions_onehot, agent_ids=IDs)
if self.use_parameter_sharing:
values_pred_dict = {k: tf.reshape(values_pred, [bs, -1]) for k in self.model_keys}
else:
values_pred_dict = {k: values_pred[:, i] for i, k in enumerate(self.model_keys)}
loss_a, loss_c = [], []
for key in self.model_keys:
mask_values = agent_mask[key]
mask_values_sum = tf.reduce_sum(mask_values)
pi_probs = tf.nn.softmax(pi_logits[key], axis=-1)
pi_probs = (1 - epsilon) * pi_probs + epsilon * 1 / self.n_actions[key]
baseline = tf.reshape(tf.reduce_sum(pi_probs * values_pred_dict[key], -1), [bs])
pi_taken = tf.gather(pi_probs, actions[key], axis=-1, batch_dims=-1)
q_taken = tf.reshape(tf.gather(values_pred_dict[key], actions[key], axis=-1, batch_dims=-1), [bs])
log_pi_taken = tf.reshape(tf.math.log(pi_taken), [bs])
advantages = tf.stop_gradient(q_taken - baseline)
loss_a.append(-tf.reduce_sum(advantages * log_pi_taken * mask_values) / mask_values_sum)
td_error = (q_taken - returns[key]) * mask_values
loss_c.append(tf.reduce_sum(td_error ** 2) / mask_values_sum)
# update critic
loss_critic = sum(loss_c)
gradients_critic = tape.gradient(loss_critic, self.policy.parameters_critic)
if self.use_grad_clip:
gradients_critic, _ = tf.clip_by_global_norm(gradients_critic, clip_norm=self.grad_clip_norm)
self.optimizer['critic'].apply_gradients(zip(gradients_critic, self.policy.parameters_critic))
else:
self.optimizer['critic'].apply_gradients(zip(gradients_critic, self.policy.parameters_critic))
# update actor
loss_coma = sum(loss_a)
gradients_actor = tape.gradient(loss_coma, self.policy.parameters_actor)
if self.use_grad_clip:
gradients_actor, _ = tf.clip_by_global_norm(gradients_actor, clip_norm=self.grad_clip_norm)
self.optimizer['actor'].apply_gradients(zip(gradients_actor, self.policy.parameters_actor))
else:
self.optimizer['actor'].apply_gradients(zip(gradients_actor, self.policy.parameters_actor))
return loss_coma, loss_critic
# @tf.function
[docs]
def learn(self, *inputs):
if self.distributed_training:
loss_coma, loss_critic = self.policy.mirrored_strategy.run(self.forward_fn, args=inputs)
return (self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss_coma, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss_critic, axis=None))
else:
return self.forward_fn(*inputs)
[docs]
def update(self, sample, epsilon=0.0):
self.iterations += 1
# prepare training data
sample_Tensor = self.build_training_data(sample=sample,
use_parameter_sharing=self.use_parameter_sharing,
use_actions_mask=self.use_actions_mask,
use_global_state=True)
batch_size = sample_Tensor['batch_size']
state = sample_Tensor['state']
obs = sample_Tensor['obs']
actions = sample_Tensor['actions']
agent_mask = sample_Tensor['agent_mask']
avail_actions = sample_Tensor['avail_actions']
returns = sample_Tensor['returns']
IDs = sample_Tensor['agent_ids']
bs = batch_size * self.n_agents if self.use_parameter_sharing else batch_size
info = self.callback.on_update_start(self.iterations, method="update",
policy=self.policy, sample_Tensor=sample_Tensor, bs=bs)
loss_coma, loss_critic = self.learn(bs, batch_size, obs, state, actions,
agent_mask, avail_actions, returns, IDs, epsilon)
if self.iterations % self.sync_frequency == 0:
self.policy.copy_target()
info.update({
"actor_loss": loss_coma.numpy(),
"critic_loss": loss_critic.numpy(),
})
info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info))
return info