"""
Multi-Agent Deep Deterministic Policy Gradient
Paper link:
https://proceedings.neurips.cc/paper/2017/file/68a9750337a418a86fe06c1991a1d64c-Paper.pdf
Implementation: Pytorch
Trick: Parameter sharing for all agents, with agents' one-hot IDs as actor-critic's inputs.
"""
import torch
from torch import nn
from xuance.torch.learners import LearnerMAS
from xuance.common import List
from argparse import Namespace
[docs]
class MADDPG_Learner(LearnerMAS):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: nn.Module,
callback):
super(MADDPG_Learner, self).__init__(config, model_keys, agent_keys, policy, callback)
self.optimizer = {
key: {'actor': torch.optim.Adam(self.policy.parameters_actor[key], self.config.learning_rate_actor, eps=1e-5),
'critic': torch.optim.Adam(self.policy.parameters_critic[key], self.config.learning_rate_critic, eps=1e-5)}
for key in self.model_keys}
self.scheduler = {
key: {'actor': torch.optim.lr_scheduler.LinearLR(self.optimizer[key]['actor'],
start_factor=1.0,
end_factor=self.end_factor_lr_decay,
total_iters=self.total_iters),
'critic': torch.optim.lr_scheduler.LinearLR(self.optimizer[key]['critic'],
start_factor=1.0,
end_factor=self.end_factor_lr_decay,
total_iters=self.total_iters)}
for key in self.model_keys}
self.gamma = config.gamma
self.tau = config.tau
self.mse_loss = nn.MSELoss()
[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 = self.get_joint_input(obs, (batch_size, -1))
next_obs_joint = self.get_joint_input(obs_next, (batch_size, -1))
actions_joint = self.get_joint_input(actions, (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)
# get actions
_, actions_eval = self.policy(observation=obs, agent_ids=IDs)
_, actions_next = self.policy.Atarget(next_observation=obs_next, agent_ids=IDs)
# get values
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 = self.get_joint_input(actions_next, (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]
# update critic
q_eval_a = q_eval[key].reshape(bs)
q_next_i = q_next[key].reshape(bs)
q_target = rewards[key] + (1 - terminals[key]) * self.gamma * q_next_i
td_error = (q_eval_a - q_target.detach()) * mask_values
loss_c = (td_error ** 2).sum() / mask_values.sum()
self.optimizer[key]['critic'].zero_grad()
loss_c.backward()
if self.use_grad_clip:
torch.nn.utils.clip_grad_norm_(self.policy.parameters_critic[key], self.grad_clip_norm)
self.optimizer[key]['critic'].step()
if self.scheduler[key]['critic'] is not None:
self.scheduler[key]['critic'].step()
# update actor
if self.use_parameter_sharing:
act_eval = actions_eval[key].reshape(batch_size, self.n_agents, -1).reshape(batch_size, -1)
else:
a_joint = {k: actions_eval[k] if k == key else actions[k] for k in self.agent_keys}
act_eval = self.get_joint_input(a_joint, (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 = q_policy[key].reshape(bs)
loss_a = -(q_policy_i * mask_values).sum() / mask_values.sum()
self.optimizer[key]['actor'].zero_grad()
loss_a.backward()
if self.use_grad_clip:
torch.nn.utils.clip_grad_norm_(self.policy.parameters_actor[key], self.grad_clip_norm)
self.optimizer[key]['actor'].step()
if self.scheduler[key]['actor'] is not None:
self.scheduler[key]['actor'].step()
learning_rate_actor = self.optimizer[key]['actor'].state_dict()['param_groups'][0]['lr']
learning_rate_critic = self.optimizer[key]['critic'].state_dict()['param_groups'][0]['lr']
info.update({
f"{key}/learning_rate_actor": learning_rate_actor,
f"{key}/learning_rate_critic": learning_rate_critic,
f"{key}/loss_actor": loss_a.item(),
f"{key}/loss_critic": loss_c.item(),
f"{key}/predictQ": q_eval[key].mean().item()
})
info.update(self.callback.on_update_agent_wise(self.iterations, key, info=info, method="update",
mask_values=mask_values, q_policy_i=q_policy_i,
act_eval=act_eval, q_eval_a=q_eval_a, q_next_i=q_next_i,
q_target=q_target, td_error=td_error))
self.policy.soft_update(self.tau)
info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info))
return info
[docs]
def update_rnn(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']
seq_len = sample_Tensor['seq_length']
obs = sample_Tensor['obs']
actions = sample_Tensor['actions']
rewards = sample_Tensor['rewards']
terminals = sample_Tensor['terminals']
agent_mask = sample_Tensor['agent_mask']
filled = sample_Tensor['filled']
IDs = sample_Tensor['agent_ids']
if self.use_parameter_sharing:
key = self.model_keys[0]
bs_rnn = batch_size * self.n_agents
filled = filled.unsqueeze(1).expand(-1, self.n_agents, -1).reshape(bs_rnn, seq_len)
obs_joint = obs[key].reshape(batch_size, self.n_agents, seq_len + 1, -1).transpose(
1, 2).reshape(batch_size, seq_len + 1, -1)
actions_joint = actions[key].reshape(batch_size, self.n_agents, seq_len, -1).transpose(
1, 2).reshape(batch_size, seq_len, -1)
rewards[key] = rewards[key].reshape(bs_rnn, seq_len)
terminals[key] = terminals[key].reshape(bs_rnn, seq_len)
IDs_t = IDs[:, :-1]
else:
bs_rnn, IDs_t = batch_size, None
obs_joint = self.get_joint_input(obs, (batch_size, seq_len + 1, -1))
actions_joint = self.get_joint_input(actions, (batch_size, seq_len, -1))
info = self.callback.on_update_start(self.iterations, method="update_rnn",
policy=self.policy, sample_Tensor=sample_Tensor, bs_rnn=bs_rnn,
obs_joint=obs_joint, actions_joint=actions_joint)
# initial hidden states for rnn
rnn_hidden_actor = {k: self.policy.actor_representation[k].init_hidden(bs_rnn) for k in self.model_keys}
rnn_hidden_critic = {k: self.policy.critic_representation[k].init_hidden(batch_size) for k in self.model_keys}
# get actions
_, actions_eval = self.policy(observation=obs, agent_ids=IDs, rnn_hidden=rnn_hidden_actor)
_, actions_next = self.policy.Atarget(next_observation=obs, agent_ids=IDs, rnn_hidden=rnn_hidden_actor)
# get q values
if self.use_parameter_sharing:
key = self.model_keys[0]
actions_next_joint = actions_next[key].reshape(batch_size, self.n_agents, seq_len + 1, -1).transpose(
1, 2).reshape(batch_size, seq_len + 1, -1)
else:
actions_next_joint = self.get_joint_input(actions_next, (batch_size, seq_len + 1, -1))
_, q_eval = self.policy.Qpolicy(joint_observation=obs_joint[:, :-1], joint_actions=actions_joint,
agent_ids=IDs_t, rnn_hidden=rnn_hidden_critic)
_, q_next = self.policy.Qtarget(joint_observation=obs_joint, joint_actions=actions_next_joint, agent_ids=IDs,
rnn_hidden=rnn_hidden_critic)
for key in self.model_keys:
mask_values = agent_mask[key] * filled
# update critic
q_eval_a = q_eval[key].reshape(bs_rnn, seq_len)
q_next_i = q_next[key][:, 1:].reshape(bs_rnn, seq_len)
q_target = rewards[key] + (1 - terminals[key]) * self.gamma * q_next_i
td_error = (q_eval_a - q_target.detach()) * mask_values
loss_c = (td_error ** 2).sum() / mask_values.sum()
self.optimizer[key]['critic'].zero_grad()
loss_c.backward()
if self.use_grad_clip:
torch.nn.utils.clip_grad_norm_(self.policy.parameters_critic[key], self.grad_clip_norm)
self.optimizer[key]['critic'].step()
if self.scheduler[key]['critic'] is not None:
self.scheduler[key]['critic'].step()
# update actor
if self.use_parameter_sharing:
act_eval = actions_eval[key][:, :-1].reshape(
batch_size, self.n_agents, seq_len, -1).transpose(1, 2).reshape(batch_size, seq_len, -1)
else:
a_dict = {k: actions_eval[k][:, :-1] if k == key else actions[k] for k in self.agent_keys}
act_eval = self.get_joint_input(a_dict, (batch_size, seq_len, -1))
_, q_policy = self.policy.Qpolicy(joint_observation=obs_joint[:, :-1], joint_actions=act_eval,
agent_key=key, agent_ids=IDs_t, rnn_hidden=rnn_hidden_critic)
q_policy_i = q_policy[key].reshape(bs_rnn, seq_len)
loss_a = -(q_policy_i * mask_values).sum() / mask_values.sum()
self.optimizer[key]['actor'].zero_grad()
loss_a.backward()
if self.use_grad_clip:
torch.nn.utils.clip_grad_norm_(self.policy.parameters_actor[key], self.grad_clip_norm)
self.optimizer[key]['actor'].step()
if self.scheduler[key]['actor'] is not None:
self.scheduler[key]['actor'].step()
learning_rate_actor = self.optimizer[key]['actor'].state_dict()['param_groups'][0]['lr']
learning_rate_critic = self.optimizer[key]['critic'].state_dict()['param_groups'][0]['lr']
info.update({
f"{key}/learning_rate_actor": learning_rate_actor,
f"{key}/learning_rate_critic": learning_rate_critic,
f"{key}/loss_actor": loss_a.item(),
f"{key}/loss_critic": loss_c.item(),
f"{key}/predictQ": q_eval[key].mean().item()
})
info.update(self.callback.on_update_agent_wise(self.iterations, key, info=info, method="update_rnn",
mask_values=mask_values, q_policy_i=q_policy_i,
q_eval_a=q_eval_a, q_next_i=q_next_i,
q_target=q_target, td_error=td_error))
self.policy.soft_update(self.tau)
info.update(self.callback.on_update_end(self.iterations, method="update_rnn", policy=self.policy, info=info))
return info