"""
Multi-Agent Proximal Policy Optimization (MAPPO)
Paper link:
https://arxiv.org/pdf/2103.01955.pdf
Implementation: MindSpore
"""
from argparse import Namespace
from xuance.common import List
from xuance.mindspore import ms, nn, msd, ops, Module, Tensor
from xuance.mindspore.utils import ValueNorm, clip_grads
from xuance.mindspore.learners.multi_agent_rl.iac_learner import IAC_Learner
[docs]
class IPPO_Learner(IAC_Learner):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
super(IPPO_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
self.lr = config.learning_rate
self.end_factor_lr_decay = config.end_factor_lr_decay
self.gamma = config.gamma
self.clip_range = config.clip_range
self.use_linear_lr_decay = config.use_linear_lr_decay
self.use_value_clip, self.value_clip_range = config.use_value_clip, config.value_clip_range
self.use_huber_loss, self.huber_delta = config.use_huber_loss, config.huber_delta
self.use_value_norm = config.use_value_norm
self.use_global_state = config.use_global_state
self.vf_coef, self.ent_coef = config.vf_coef, config.ent_coef
self.mse_loss = nn.MSELoss()
self.huber_loss = nn.HuberLoss(reduction="none", delta=self.huber_delta)
self.softmax = nn.Softmax(axis=-1)
self.is_continuous = self.policy.is_continuous
if self.use_value_norm:
self.value_normalizer = {key: ValueNorm(1) for key in self.model_keys}
else:
self.value_normalizer = None
if self.is_continuous:
self.pi_dist = {k: msd.Normal(dtype=ms.float32) for k in self.model_keys}
else:
self.pi_dist = {k: msd.Categorical() for k in self.model_keys}
# Get gradient function
self.grad_fn = ms.value_and_grad(self.forward_fn, None, self.optimizer.parameters, has_aux=True)
self.policy.set_train()
[docs]
def forward_fn(self, *args):
bs, obs, actions, avail_actions, log_pi_old, values, returns, advantages, agt_mask, ids = args
info_forward = {}
pi_dist_mu, pi_dist_std, pi_dist_logits = {}, {}, {}
# feedforward
if self.is_continuous:
_, pi_dist_mu, pi_dist_std = self.policy(observation=obs, agent_ids=ids, avail_actions=avail_actions)
else:
_, pi_dist_logits = self.policy(observation=obs, agent_ids=ids, avail_actions=avail_actions)
_, value_pred_dict = self.policy.get_values(observation=obs, agent_ids=ids)
# calculate losses for each agent
loss_a, loss_e, loss_c = [], [], []
for key in self.model_keys:
mask_values = agt_mask[key]
# actor loss
if self.is_continuous:
log_pi = self.pi_dist[key]._log_prob(value=actions[key], mean=pi_dist_mu[key], sd=pi_dist_std[key])
log_pi = ops.reduce_sum(x=log_pi, axis=-1)
entropy = self.pi_dist[key]._entropy(mean=pi_dist_mu[key], sd=pi_dist_std[key])
entropy = ops.reduce_sum(x=entropy, axis=-1)
else:
probs = self.softmax(pi_dist_logits[key])
log_pi = self.pi_dist[key]._log_prob(value=actions[key], probs=probs)
entropy = self.pi_dist[key].entropy(probs=probs)
ratio = ops.exp(log_pi - log_pi_old[key]).reshape(bs)
advantages_mask = ops.stop_gradient(advantages[key]) * mask_values
surrogate1 = ratio * advantages_mask
surrogate2 = ops.clip_by_value(ratio, Tensor(1 - self.clip_range), Tensor(1 + self.clip_range))
surrogate2 *= advantages_mask
loss_a.append(-ops.minimum(surrogate1, surrogate2).mean())
# entropy loss
entropy_loss = (entropy * mask_values).sum() / mask_values.sum()
loss_e.append(entropy_loss)
# critic loss
value_pred_i = value_pred_dict[key].reshape(bs)
value_target = returns[key].reshape(bs)
values_i = values[key].reshape(bs)
if self.use_value_clip:
value_clipped = values_i + (value_pred_i - values_i).clamp(-self.value_clip_range,
self.value_clip_range)
if self.use_value_norm:
self.value_normalizer[key].update(value_target.reshape(bs, 1))
value_target = self.value_normalizer[key].normalize(value_target.reshape(bs, 1)).reshape(bs)
if self.use_huber_loss:
loss_v = self.huber_loss(value_pred_i, value_target)
loss_v_clipped = self.huber_loss(value_clipped, value_target)
else:
loss_v = (value_pred_i - value_target) ** 2
loss_v_clipped = (value_clipped - value_target) ** 2
loss_c_ = ops.maximum(loss_v, loss_v_clipped) * mask_values
loss_c.append(loss_c_.sum() / mask_values.sum())
else:
if self.use_value_norm:
self.value_normalizer[key].update(value_target)
value_target = self.value_normalizer[key].normalize(value_target)
if self.use_huber_loss:
loss_v = self.huber_loss(value_pred_i, value_target) * mask_values
else:
loss_v = ((value_pred_i - value_target) ** 2) * mask_values
loss_c.append(loss_v.sum() / mask_values.sum())
info_forward.update({
f"predict_value/{key}": value_pred_i.mean().asnumpy()
})
info_forward.update(self.callback.on_update_agent_wise(self.iterations, key, info=info_forward,
method="update",
mask_values=mask_values, log_pi=log_pi, ratio=ratio,
surrogate1=surrogate1, surrogate2=surrogate2,
entropy=entropy,
value_pred_i=value_pred_i, value_target=value_target,
values_i=values_i, loss_v=loss_v))
loss = sum(loss_a) + self.vf_coef * sum(loss_c) - self.ent_coef * sum(loss_e)
return loss, sum(loss_a), sum(loss_c), sum(loss_e), info_forward
[docs]
def update(self, sample):
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)
batch_size = sample_Tensor['batch_size']
obs = sample_Tensor['obs']
actions = sample_Tensor['actions']
agent_mask = sample_Tensor['agent_mask']
avail_actions = sample_Tensor['avail_actions']
values = sample_Tensor['values']
returns = sample_Tensor['returns']
advantages = sample_Tensor['advantages']
log_pi_old = sample_Tensor['log_pi_old']
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, loss_a, loss_c, loss_e, info_forward), grads = self.grad_fn(bs, obs, actions, avail_actions, log_pi_old,
values, returns, advantages, agent_mask, IDs)
if self.use_grad_clip:
grads = clip_grads(grads, Tensor(-self.grad_clip_norm), Tensor(self.grad_clip_norm))
self.optimizer(grads)
self.scheduler.step()
lr = self.scheduler.get_last_lr()[0]
info.update({
"learning_rate": lr.asnumpy(),
"loss": loss.asnumpy(),
"loss_a": loss_a.asnumpy(),
"loss_c": loss_c.asnumpy(),
"loss_e": loss_e.asnumpy()
})
info.update(info_forward)
info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info))
return info