Source code for xuance.torch.learners.multi_agent_rl.wqmix_learner

"""
Weighted QMIX
Paper link: https://proceedings.neurips.cc/paper/2020/file/73a427badebe0e32caa2e1fc7530b7f3-Paper.pdf
Implementation: Pytorch
"""
import torch
from torch import nn
from xuance.torch.learners import LearnerMAS
from xuance.common import List
from argparse import Namespace
from operator import itemgetter


[docs] class WQMIX_Learner(LearnerMAS): def __init__(self, config: Namespace, model_keys: List[str], agent_keys: List[str], policy: nn.Module, callback): super(WQMIX_Learner, self).__init__(config, model_keys, agent_keys, policy, callback) self.optimizer = torch.optim.Adam(self.policy.parameters_model, config.learning_rate, eps=1e-5) self.scheduler = torch.optim.lr_scheduler.LinearLR(self.optimizer, start_factor=1.0, end_factor=self.end_factor_lr_decay, total_iters=self.total_iters) self.alpha = config.alpha self.sync_frequency = config.sync_frequency self.mse_loss = nn.MSELoss() self.n_actions = {k: self.policy.action_space[k].n for k in self.model_keys}
[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=True) batch_size = sample_Tensor['batch_size'] state = sample_Tensor['state'] state_next = sample_Tensor['state_next'] 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'] avail_actions = sample_Tensor['avail_actions'] avail_actions_next = sample_Tensor['avail_actions_next'] IDs = sample_Tensor['agent_ids'] if self.use_parameter_sharing: key = self.model_keys[0] bs = batch_size * self.n_agents rewards_tot = rewards[key].mean(dim=1).reshape(batch_size, 1) terminals_tot = terminals[key].all(dim=1, keepdim=False).float().reshape(batch_size, 1) else: bs = batch_size rewards_tot = torch.stack(itemgetter(*self.agent_keys)(rewards), dim=1).mean(dim=-1, keepdim=True) terminals_tot = torch.stack(itemgetter(*self.agent_keys)(terminals), dim=1).all(dim=1, keepdim=True).float() info = self.callback.on_update_start(self.iterations, method="update", policy=self.policy, sample_Tensor=sample_Tensor, bs=bs, rewards_tot=rewards_tot, terminals_tot=terminals_tot) # calculate Q_tot _, action_max, q_eval = self.policy(observation=obs, agent_ids=IDs, avail_actions=avail_actions) _, q_eval_centralized = self.policy.q_centralized(observation=obs, agent_ids=IDs) _, q_eval_next_centralized = self.policy.target_q_centralized(observation=obs_next, agent_ids=IDs) q_eval_a, q_eval_centralized_a, q_eval_next_centralized_a, act_next = {}, {}, {}, {} for key in self.model_keys: mask_values = agent_mask[key] action_max[key] = action_max[key].unsqueeze(-1) q_eval_a[key] = q_eval[key].gather(-1, actions[key].long().unsqueeze(-1)).reshape(bs) q_eval_centralized_a[key] = q_eval_centralized[key].gather(-1, action_max[key].long()).reshape(bs) if self.config.double_q: _, a_next_greedy, _ = self.policy(observation=obs_next, agent_ids=IDs, avail_actions=avail_actions_next, agent_key=key) act_next[key] = a_next_greedy[key].unsqueeze(-1) else: _, q_next_eval = self.policy.Qtarget(observation=obs_next, agent_ids=IDs, agent_key=key) if self.use_actions_mask: q_next_eval[key][avail_actions_next[key] == 0] = -1e10 act_next[key] = q_next_eval[key].argmax(dim=-1, keepdim=True) q_eval_next_centralized_a[key] = q_eval_next_centralized[key].gather(-1, act_next[key]).reshape(bs) q_eval_a[key] *= mask_values q_eval_centralized_a[key] *= mask_values q_eval_next_centralized_a[key] *= mask_values info.update(self.callback.on_update_agent_wise(self.iterations, key, info=info, method="update", mask_values=mask_values, q_eval_a=q_eval_a, q_eval_centralized_a=q_eval_centralized_a, act_next=act_next, q_eval_next_centralized_a=q_eval_next_centralized_a)) q_tot_eval = self.policy.Q_tot(q_eval_a, state) # calculate Q_tot q_tot_centralized = self.policy.q_feedforward(q_eval_centralized_a, state) # calculate centralized Q q_tot_next_centralized = self.policy.target_q_feedforward(q_eval_next_centralized_a, state_next) # y_i target_value = rewards_tot + (1 - terminals_tot) * self.gamma * q_tot_next_centralized td_error = q_tot_eval - target_value.detach() # calculate weights ones = torch.ones_like(td_error) w = ones * self.alpha if self.config.agent == "CWQMIX": condition_1 = ((action_max == actions.reshape([-1, self.n_agents, 1])) * agent_mask).all(dim=1) condition_2 = target_value > q_tot_centralized conditions = condition_1 | condition_2 w = torch.where(conditions, ones, w) elif self.config.agent == "OWQMIX": condition = td_error < 0 w = torch.where(condition, ones, w) else: raise AttributeError(f"The agent named is {self.config.agent} is currently not supported.") # calculate losses and train loss_central = self.mse_loss(q_tot_centralized, target_value.detach()) loss_qmix = (w.detach() * (td_error ** 2)).mean() loss = loss_qmix + loss_central self.optimizer.zero_grad() loss.backward() if self.use_grad_clip: torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.grad_clip_norm) self.optimizer.step() if self.scheduler is not None: self.scheduler.step() if self.iterations % self.sync_frequency == 0: self.policy.copy_target() lr = self.optimizer.state_dict()['param_groups'][0]['lr'] info.update({ "learning_rate": lr, "loss_Qmix": loss_qmix.item(), "loss_central": loss_central.item(), "loss": loss.item(), "predictQ": q_tot_eval.mean().item() }) info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info, q_tot_eval=q_tot_eval, q_tot_centralized=q_tot_centralized, q_tot_next_centralized=q_tot_next_centralized, target_value=target_value, td_error=td_error, ones=ones, w=w)) 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, use_global_state=True) batch_size = sample_Tensor['batch_size'] seq_len = sample['sequence_length'] state = sample_Tensor['state'] obs = sample_Tensor['obs'] actions = sample_Tensor['actions'] rewards = sample_Tensor['rewards'] terminals = sample_Tensor['terminals'] agent_mask = sample_Tensor['agent_mask'] avail_actions = sample_Tensor['avail_actions'] filled = sample_Tensor['filled'].reshape([-1, 1]) IDs = sample_Tensor['agent_ids'] if self.use_parameter_sharing: key = self.model_keys[0] bs_rnn = batch_size * self.n_agents rewards_tot = rewards[key].mean(dim=1).reshape([-1, 1]) terminals_tot = terminals[key].all(dim=1, keepdim=False).float().reshape([-1, 1]) else: bs_rnn = batch_size rewards_tot = torch.stack(itemgetter(*self.agent_keys)(rewards), dim=1).mean(dim=1).reshape(-1, 1) terminals_tot = torch.stack(itemgetter(*self.agent_keys)(terminals), dim=1).all(1).reshape([-1, 1]).float() info = self.callback.on_update_start(self.iterations, method="update_rnn", policy=self.policy, sample_Tensor=sample_Tensor, bs_rnn=bs_rnn, rewards_tot=rewards_tot, terminals_tot=terminals_tot) # calculate Q_tot rnn_hidden = {k: self.policy.representation[k].init_hidden(bs_rnn) for k in self.model_keys} _, action_max, q_eval = self.policy(observation=obs, agent_ids=IDs, avail_actions=avail_actions, rnn_hidden=rnn_hidden) rnn_hidden_cent = {k: self.policy.representation[k].init_hidden(bs_rnn) for k in self.model_keys} _, q_eval_centralized = self.policy.q_centralized(observation=obs, agent_ids=IDs, rnn_hidden=rnn_hidden_cent) target_rnn_hidden_cent = {k: self.policy.target_representation[k].init_hidden(bs_rnn) for k in self.model_keys} _, q_eval_next_centralized = self.policy.target_q_centralized(observation=obs, agent_ids=IDs, rnn_hidden=target_rnn_hidden_cent) q_eval_a, q_eval_centralized_a, q_eval_next_centralized_a = {}, {}, {} target_rnn_hidden = {k: self.policy.target_representation[k].init_hidden(bs_rnn) for k in self.model_keys} for key in self.model_keys: mask_values = agent_mask[key] act_greedy = action_max[key][:, :-1].unsqueeze(-1) q_eval_a[key] = q_eval[key][:, :-1].gather(-1, actions[key].long().unsqueeze(-1)).reshape(bs_rnn, seq_len) q_eval_centralized_a[key] = q_eval_centralized[key][:, :-1].gather(-1, act_greedy.long()).reshape(bs_rnn, seq_len) if self.config.double_q: act_next = action_max[key][:, 1:].unsqueeze(-1) else: _, q_next_seq = self.policy.Qtarget(observation=obs, agent_ids=IDs, agent_key=key, rnn_hidden=target_rnn_hidden) q_next_eval = q_next_seq[key][:, 1:] if self.use_actions_mask: q_next_eval[avail_actions[key][:, 1:] == 0] = -1e10 act_next = q_next_eval.argmax(dim=-1, keepdim=True) q_eval_next_centralized_a[key] = q_eval_next_centralized[key][:, 1:].gather(-1, act_next).reshape(bs_rnn, seq_len) q_eval_a[key] *= mask_values q_eval_centralized_a[key] *= mask_values q_eval_next_centralized_a[key] *= mask_values if self.use_parameter_sharing: q_eval_a[key] = q_eval_a[key].reshape(batch_size, self.n_agents, seq_len).transpose(1, 2).reshape(-1, self.n_agents) q_eval_centralized_a[key] = q_eval_centralized_a[key].reshape(batch_size, self.n_agents, seq_len).transpose(1, 2).reshape(-1, self.n_agents) q_eval_next_centralized_a[key] = q_eval_next_centralized_a[key].reshape(batch_size, self.n_agents, seq_len).transpose(1, 2).reshape(-1, self.n_agents) else: q_eval_a[key] = q_eval_a[key].reshape(-1, 1) q_eval_centralized_a[key] = q_eval_centralized_a[key].reshape(-1, 1) q_eval_next_centralized_a[key] = q_eval_next_centralized_a[key].reshape(-1, 1) info.update(self.callback.on_update_agent_wise(self.iterations, key, info=info, method="update_rnn", mask_values=mask_values, q_eval_a=q_eval_a, q_eval_centralized_a=q_eval_centralized_a, act_next=act_next, q_eval_next_centralized_a=q_eval_next_centralized_a)) state_input = state[:, :-1].reshape([batch_size * seq_len, -1]) state_input_next = state[:, 1:].reshape([batch_size * seq_len, -1]) q_tot_eval = self.policy.Q_tot(q_eval_a, state_input) # calculate Q_tot q_tot_centralized = self.policy.q_feedforward(q_eval_centralized_a, state_input) # calculate centralized Q q_tot_next_centralized = self.policy.target_q_feedforward(q_eval_next_centralized_a, state_input_next) # y_i target_value = rewards_tot + (1 - terminals_tot) * self.gamma * q_tot_next_centralized td_error = q_tot_eval - target_value.detach() # calculate weights ones = torch.ones_like(td_error) w = ones * self.alpha if self.config.agent == "CWQMIX": condition_1 = ((action_max == actions.reshape([-1, self.n_agents, 1])) * agent_mask).all(dim=1) condition_2 = target_value > q_tot_centralized conditions = condition_1 | condition_2 w = torch.where(conditions, ones, w) elif self.config.agent == "OWQMIX": condition = td_error < 0 w = torch.where(condition, ones, w) else: raise AttributeError(f"The agent named is {self.config.agent} is currently not supported.") # calculate losses and train loss_central = (((q_tot_centralized - target_value.detach()) ** 2) * filled).sum() / filled.sum() loss_qmix = (w.detach() * (td_error ** 2) * filled).sum() / filled.sum() loss = loss_qmix + loss_central self.optimizer.zero_grad() loss.backward() if self.use_grad_clip: torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.grad_clip_norm) self.optimizer.step() if self.scheduler is not None: self.scheduler.step() if self.iterations % self.sync_frequency == 0: self.policy.copy_target() lr = self.optimizer.state_dict()['param_groups'][0]['lr'] info.update({ "learning_rate": lr, "loss_Qmix": loss_qmix.item(), "loss_central": loss_central.item(), "loss": loss.item(), "predictQ": q_tot_eval.mean().item() }) info.update(self.callback.on_update_end(self.iterations, method="update_rnn", policy=self.policy, info=info, q_tot_eval=q_tot_eval, q_tot_centralized=q_tot_centralized, q_tot_next_centralized=q_tot_next_centralized, target_value=target_value, td_error=td_error, ones=ones, w=w)) return info