"""
Multi-agent Soft Actor-critic (MASAC)
Implementation: MindSpore
"""
from mindspore.nn import MSELoss
from xuance.mindspore import ms, Module, Tensor, optim, ops
from xuance.mindspore.learners import LearnerMAS
from xuance.mindspore.utils import clip_grads
from xuance.common import List
from argparse import Namespace
from operator import itemgetter
[docs]
class MASAC_Learner(LearnerMAS):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
super(MASAC_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
self.optimizer = {
key: {
'actor': optim.Adam(params=self.policy.parameters_actor[key], lr=self.config.learning_rate_actor,
eps=1e-5),
'critic': optim.Adam(params=self.policy.parameters_critic[key], lr=self.config.learning_rate_critic,
eps=1e-5)}
for key in self.model_keys}
self.scheduler = {
key: {'actor': optim.lr_scheduler.LinearLR(self.optimizer[key]['actor'], start_factor=1.0,
end_factor=self.end_factor_lr_decay, total_iters=self.config.running_steps),
'critic': optim.lr_scheduler.LinearLR(self.optimizer[key]['critic'], start_factor=1.0,
end_factor=self.end_factor_lr_decay, total_iters=self.config.running_steps)}
for key in self.model_keys}
self.gamma = config.gamma
self.tau = config.tau
self.alpha = {key: config.alpha for key in self.model_keys}
self.mse_loss = MSELoss()
self._ones = ops.Ones()
self.use_automatic_entropy_tuning = config.use_automatic_entropy_tuning
if self.use_automatic_entropy_tuning:
self.target_entropy = {key: -policy.action_space[key].shape[-1] for key in self.model_keys}
self.log_alpha = {key: ms.Parameter(self._ones(1, ms.float32)) for key in self.model_keys}
self.alpha = {key: ops.exp(self.log_alpha[key]) for key in self.model_keys}
self.alpha_optimizer = {key: optim.Adam(params=[self.log_alpha[key]], lr=config.learning_rate_actor)
for key in self.model_keys}
# Get gradient function
self.grad_fn_alpha = {key: ms.value_and_grad(self.forward_fn_alpha, None,
self.alpha_optimizer[key].parameters, has_aux=True)
for key in self.model_keys}
# Get gradient function
self.grad_fn_actor = {key: ms.value_and_grad(self.forward_fn_actor, None,
self.optimizer[key]['actor'].parameters, has_aux=True)
for key in self.model_keys}
self.grad_fn_critic = {key: ms.value_and_grad(self.forward_fn_critic, None,
self.optimizer[key]['critic'].parameters, has_aux=True)
for key in self.model_keys}
[docs]
def forward_fn_alpha(self, log_pi_eval_i, key):
alpha_loss = -(self.log_alpha[key] * ops.stop_gradient((log_pi_eval_i + self.target_entropy[key]))).mean()
return alpha_loss, self.log_alpha[key]
[docs]
def forward_fn_actor(self, batch_size, obs, obs_joint, ids, mask_values, agent_key):
_, actions_eval, log_pi_eval = self.policy(observation=obs, agent_ids=ids)
if self.use_parameter_sharing:
actions_eval_joint = actions_eval[agent_key].reshape(batch_size, self.n_agents, -1).reshape(batch_size, -1)
else:
actions_eval_detach_others = {k: actions_eval[k] if k == agent_key else ops.stop_gradient(actions_eval[k])
for k in self.model_keys}
actions_eval_joint = ops.cat(itemgetter(*self.model_keys)(actions_eval_detach_others),
axis=-1).reshape(batch_size, -1)
_, _, policy_q_1, policy_q_2 = self.policy.Qpolicy(joint_observation=obs_joint,
joint_actions=actions_eval_joint,
agent_ids=ids, agent_key=agent_key)
log_pi_eval_i = log_pi_eval[agent_key].reshape(-1)
policy_q = ops.minimum(policy_q_1[agent_key], policy_q_2[agent_key]).reshape(-1)
loss_a = ((self.alpha[agent_key] * log_pi_eval_i - policy_q) * mask_values).sum() / mask_values.sum()
return loss_a, log_pi_eval[agent_key], policy_q, policy_q_1, policy_q_2
[docs]
def forward_fn_critic(self, obs_joint, actions_joint, ids, mask_values, backup, agent_key):
_, _, action_q_1, action_q_2 = self.policy.Qpolicy(joint_observation=obs_joint, joint_actions=actions_joint,
agent_ids=ids)
action_q_1_i = action_q_1[agent_key].reshape(-1)
action_q_2_i = action_q_2[agent_key].reshape(-1)
td_error_1, td_error_2 = action_q_1_i - ops.stop_gradient(backup), action_q_2_i - ops.stop_gradient(backup)
td_error_1 *= mask_values
td_error_2 *= mask_values
loss_c = ((td_error_1 ** 2).sum() + (td_error_2 ** 2).sum()) / mask_values.sum()
return loss_c, action_q_1_i, action_q_2_i, td_error_1, td_error_2
[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 = obs[key].reshape(batch_size, -1)
next_obs_joint = obs_next[key].reshape(batch_size, -1)
actions_joint = actions[key].reshape(batch_size, -1)
rewards[key] = rewards[key].reshape(batch_size * self.n_agents)
terminals[key] = terminals[key].reshape(batch_size * self.n_agents)
else:
bs = batch_size
obs_joint = ops.cat(itemgetter(*self.agent_keys)(obs), axis=-1).reshape(batch_size, -1)
next_obs_joint = ops.cat(itemgetter(*self.agent_keys)(obs_next), axis=-1).reshape(batch_size, -1)
actions_joint = ops.cat(itemgetter(*self.agent_keys)(actions), axis=-1).reshape(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)
# train the model
_, actions_next, log_pi_next = self.policy(observation=obs_next, agent_ids=IDs)
if self.use_parameter_sharing:
key = self.model_keys[0]
actions_next_joint = actions_next[key].reshape(batch_size, self.n_agents, -1).reshape(batch_size, -1)
else:
actions_next_joint = ops.cat(itemgetter(*self.model_keys)(actions_next), -1).reshape(batch_size, -1)
_, _, target_q = 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]
# critic update
log_pi_next_eval = log_pi_next[key].reshape(bs)
target_value = target_q[key].reshape(bs) - self.alpha[key] * log_pi_next_eval
backup = rewards[key] + (1 - terminals[key]) * self.gamma * target_value
(loss_c, action_q_1_i, action_q_2_i, td_error_1, td_error_2), grads_critic = self.grad_fn_critic[key](
obs_joint, actions_joint, IDs, mask_values, backup, key)
if self.use_grad_clip:
grads_critic = clip_grads(grads_critic, Tensor(-self.grad_clip_norm), Tensor(self.grad_clip_norm))
self.optimizer[key]['critic'](grads_critic)
# update actor
(loss_a, log_pi_eval_i, policy_q, policy_q_1, policy_q_2), grads_actor = self.grad_fn_actor[key](
batch_size, obs, obs_joint, IDs, mask_values, key)
if self.use_grad_clip:
grads_actor = clip_grads(grads_actor, Tensor(-self.grad_clip_norm), Tensor(self.grad_clip_norm))
self.optimizer[key]['actor'](grads_actor)
# automatic entropy tuning
if self.use_automatic_entropy_tuning:
(alpha_loss, _), grads_alpha = self.grad_fn_alpha[key](log_pi_eval_i, key)
self.alpha_optimizer[key](grads_alpha)
self.alpha[key] = ops.exp(self.log_alpha[key])
else:
alpha_loss = 0
learning_rate_actor = self.scheduler[key]['actor'].get_last_lr()[0]
learning_rate_critic = self.scheduler[key]['critic'].get_last_lr()[0]
info.update({
f"{key}/learning_rate_actor": learning_rate_actor.asnumpy(),
f"{key}/learning_rate_critic": learning_rate_critic.asnumpy(),
f"{key}/loss_actor": loss_a.asnumpy(),
f"{key}/loss_critic": loss_c.asnumpy(),
f"{key}/predictQ": policy_q.mean().asnumpy(),
f"{key}/alpha_loss": alpha_loss.asnumpy(),
f"{key}/alpha": self.alpha[key].asnumpy(),
})
info.update(self.callback.on_update_agent_wise(self.iterations, key, info=info, method="update",
mask_values=mask_values,
action_q_1_i=action_q_1_i, action_q_2_i=action_q_2_i,
log_pi_next_eval=log_pi_next_eval,
target_value=target_value, backup=backup,
td_error_1=td_error_1, td_error_2=td_error_2,
policy_q_1=policy_q_1, policy_q_2=policy_q_2,
log_pi_eval_i=log_pi_eval_i, policy_q=policy_q))
self.policy.soft_update(self.tau)
info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info))
return info