"""
Multi-Agent Proximal Policy Optimization (MAPPO)
Paper link:
https://arxiv.org/pdf/2103.01955.pdf
Implementation: TensorFlow 2.X
"""
import numpy as np
from argparse import Namespace
from operator import itemgetter
from xuance.common import List
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners.multi_agent_rl.ippo_learner import IPPO_Learner
[docs]
class MAPPO_Learner(IPPO_Learner):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
super(MAPPO_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
# @tf.function
[docs]
def forward_fn(self, *args):
bs, obs, critic_input, actions, log_pi_old, agent_mask, avail_actions, values, returns, advantages, IDs = args
with tf.GradientTape() as tape:
loss_a, loss_e, loss_c = [], [], []
if self.is_continuous:
_, pi_mu, pi_std = self.policy(observation=obs, agent_ids=IDs, avail_actions=avail_actions)
for key in self.model_keys:
mask_values = agent_mask[key]
mask_values_sum = tf.reduce_sum(mask_values)
log_2pi = tf.math.log(2.0 * np.pi)
# policy gradient loss
log_std = tf.math.log(pi_std[key] + 1e-8)
log_prob = -0.5 * (
((actions[key] - pi_mu[key]) / (pi_std[key] + 1e-8)) ** 2 + 2.0 * log_std + log_2pi)
log_pi = tf.reduce_sum(log_prob, axis=-1, keepdims=False)
ratio = tf.exp(log_pi - log_pi_old[key])
advantages_mask = advantages[key] * mask_values
surrogate1 = ratio * advantages_mask
surrogate2 = tf.clip_by_value(ratio, 1 - self.clip_range, 1 + self.clip_range) * advantages_mask
loss_a.append(-tf.reduce_mean(tf.minimum(surrogate1, surrogate2)))
# entropy loss
entropy = tf.reduce_sum(0.5 + 0.5 * log_2pi + log_std, axis=-1, keepdims=True)
entropy_loss = tf.reduce_sum(entropy * mask_values) / mask_values_sum
loss_e.append(entropy_loss)
else:
_, pi_logits = self.policy(observation=obs, agent_ids=IDs, avail_actions=avail_actions)
for key in self.model_keys:
mask_values = agent_mask[key]
mask_values_sum = tf.reduce_sum(mask_values)
# policy gradient loss
log_prob = tf.nn.log_softmax(pi_logits[key], axis=-1)
log_pi = tf.gather(log_prob, actions[key], axis=-1, batch_dims=-1)
log_pi = tf.squeeze(log_pi, axis=-1)
ratio = tf.exp(log_pi - log_pi_old[key])
advantages_mask = advantages[key] * mask_values
surrogate1 = ratio * advantages_mask
surrogate2 = tf.clip_by_value(ratio, 1 - self.clip_range, 1 + self.clip_range) * advantages_mask
loss_a.append(-tf.reduce_mean(tf.minimum(surrogate1, surrogate2)))
# entropy loss
probs = tf.exp(log_prob)
entropy = -tf.reduce_sum(probs * log_prob, axis=-1, keepdims=False)
entropy_loss = tf.reduce_sum(entropy * mask_values) / mask_values_sum
loss_e.append(entropy_loss)
_, values_pred_dict = self.policy.get_values(observation=critic_input, agent_ids=IDs)
for key in self.model_keys:
# value loss
value_pred_i = tf.reshape(values_pred_dict[key], [bs])
value_target = tf.reshape(returns[key], [bs])
values_i = tf.reshape(values[key], [bs])
if self.use_value_clip:
value_clipped = values_i + tf.clip_by_value(value_pred_i - values_i,
-self.value_clip_range, self.value_clip_range)
if self.use_value_norm:
self.value_normalizer[key].update(tf.reshape(value_target, [bs, 1]))
value_target = tf.reshape(self.value_normalizer[key].normalize(tf.reshape(value_target,
[bs, 1])), [bs])
if self.use_huber_loss:
loss_v = tk.losses.huber(value_target, value_pred_i, self.huber_delta)
loss_v_clipped = tk.losses.huber(value_target, value_clipped, self.huber_delta)
else:
loss_v = (value_pred_i - value_target) ** 2
loss_v_clipped = (value_clipped - value_target) ** 2
loss_c_ = tf.maximum(loss_v, loss_v_clipped) * mask_values
loss_c.append(tf.reduce_sum(loss_c_) / 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 = tk.losses.huber(value_target, value_pred_i, self.huber_delta) * mask_values
else:
loss_v = ((value_pred_i - value_target) ** 2) * mask_values
loss_c.append(tf.reduce_sum(loss_v) / mask_values_sum)
# Total loss
loss = sum(loss_a) + self.vf_coef * sum(loss_c) - self.ent_coef * sum(loss_e)
gradients = tape.gradient(loss, self.policy.trainable_variables)
if self.use_grad_clip:
gradients, _ = tf.clip_by_global_norm(gradients, clip_norm=self.grad_clip_norm)
self.optimizer.apply_gradients(zip(gradients, self.policy.trainable_variables))
else:
self.optimizer.apply_gradients(zip(gradients, self.policy.trainable_variables))
return loss, loss_a, loss_c, loss_e, values_pred_dict
# @tf.function
[docs]
def learn(self, *inputs):
if self.distributed_training:
loss, a_loss, c_loss, e_loss, v_pred = self.policy.mirrored_strategy.run(self.forward_fn, args=inputs)
return (self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, a_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, c_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, e_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, v_pred, axis=None))
else:
return self.forward_fn(*inputs)
[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,
use_global_state=self.use_global_state)
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']
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']
# prepare critic inputs
if self.use_parameter_sharing:
key = self.model_keys[0]
bs = batch_size * self.n_agents
if self.use_global_state:
critic_input = {key: tf.reshape(tf.tile(tf.reshape(state, [batch_size, 1, -1]),
[1, self.n_agents, 1]), [bs, -1])}
else:
critic_input = {key: tf.reshape(tf.tile(tf.reshape(obs[key], [batch_size, 1, -1]),
[1, self.n_agents, 1]), [bs, -1])}
else:
bs = batch_size
if self.use_global_state:
critic_input = {k: tf.reshape(state, [batch_size, -1]) for k in self.agent_keys}
else:
if self.n_agents == 1:
joint_obs = itemgetter(*self.agent_keys)(obs)
else:
joint_obs = tf.concat(itemgetter(*self.agent_keys)(obs), axis=-1)
critic_input = {k: joint_obs for k in self.agent_keys}
info = self.callback.on_update_start(self.iterations, method="update",
policy=self.policy, sample_Tensor=sample_Tensor, bs=bs,
critic_input=critic_input)
loss, a_loss, c_loss, e_loss, v_pred = self.learn(bs, obs, critic_input, actions, log_pi_old,
agent_mask, avail_actions, values, returns, advantages, IDs)
info.update({f"predict_value/{key}": tf.reduce_mean(v_pred[key]).numpy() for key in self.model_keys})
info.update({
# "learning_rate": lr,
"pg_loss": sum(a_loss).numpy(),
"vf_loss": sum(c_loss).numpy(),
"entropy_loss": sum(e_loss).numpy(),
"loss": loss.numpy(),
})
info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info))
return info