Source code for xuance.tensorflow.learners.multi_agent_rl.qtran_learner

"""
QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning
Paper link:
http://proceedings.mlr.press/v97/son19a/son19a.pdf
Implementation: TensorFlow 2.X
"""
from argparse import Namespace
from operator import itemgetter
from xuance.common import List
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners import LearnerMAS


[docs] class QTRAN_Learner(LearnerMAS): def __init__(self, config: Namespace, model_keys: List[str], agent_keys: List[str], policy: Module, callback): super(QTRAN_Learner, self).__init__(config, model_keys, agent_keys, policy, callback) self.build_optimizer() self.n_actions = {k: self.policy.action_space[k].n for k in self.model_keys} self.sync_frequency = config.sync_frequency self.mse_loss = tk.losses.MeanSquaredError()
[docs] def build_optimizer(self): if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips. self.optimizer = tk.optimizers.legacy.Adam(self.config.learning_rate) else: self.optimizer = tk.optimizers.Adam(self.config.learning_rate)
# @tf.function
[docs] def forward_fn(self, bs, batch_size, state, obs, actions, rewards_tot, state_next, obs_next, terminals_tot, agent_mask, avail_actions, avail_actions_next, IDs): with tf.GradientTape() as tape: _, hidden_state, actions_greedy, q_eval = self.policy(obs, agent_ids=IDs, avail_actions=avail_actions) _, hidden_state_next, q_next = self.policy.Qtarget(obs_next, agent_ids=IDs) q_eval_a, q_eval_greedy_a, q_next_a = {}, {}, {} actions_next_greedy = {} for key in self.model_keys: mask_values = agent_mask[key] q_eval_a[key] = tf.reshape(tf.gather(q_eval[key], tf.cast(actions[key][:, None], dtype=tf.int32), axis=-1, batch_dims=-1), [bs]) q_eval_greedy_a[key] = tf.reshape(tf.gather(q_eval[key], tf.cast(actions_greedy[key][:, None], dtype=tf.int32), axis=-1, batch_dims=-1), [bs]) if self.use_actions_mask: q_next[key][avail_actions_next[key] == 0] = -1e10 if self.config.double_q: _, _, act_next, _ = self.policy(observation=obs_next, agent_ids=IDs, avail_actions=avail_actions, agent_key=key) actions_next_greedy[key] = act_next[key] q_next_a[key] = tf.reshape(tf.gather(q_next[key], act_next[key][:, None], axis=-1, batch_dims=-1), [bs]) else: actions_next_greedy[key] = q_next[key].argmax(dim=-1, keepdim=False) q_next_a[key] = q_next[key].max(dim=-1, keepdim=True).values.reshape(bs) q_eval_a[key] *= mask_values q_eval_greedy_a[key] *= mask_values q_next_a[key] *= mask_values if self.config.agent == "QTRAN_base": # -- TD Loss -- q_joint, v_joint = self.policy.Q_tran(state, hidden_state, actions, agent_mask) q_joint_next, _ = self.policy.Q_tran_target(state_next, hidden_state_next, actions_next_greedy, agent_mask) y_dqn = rewards_tot + (1 - terminals_tot) * self.gamma * q_joint_next loss_td = self.mse_loss(tf.stop_gradient(y_dqn), q_joint) # TD loss # -- Opt Loss -- # Argmax across the current agents' actions q_tot_greedy = self.policy.Q_tot(q_eval_greedy_a) q_joint_greedy_hat, _ = self.policy.Q_tran(state, hidden_state, actions_greedy, agent_mask) error_opt = q_tot_greedy - tf.stop_gradient(q_joint_greedy_hat) + v_joint loss_opt = tf.reduce_mean(error_opt ** 2) # Opt loss # -- Nopt Loss -- q_tot = self.policy.Q_tot(q_eval_a) q_joint_hat = q_joint error_nopt = q_tot - tf.stop_gradient(q_joint_hat) + v_joint error_nopt = tf.clip_by_value(error_nopt, clip_value_min=-1e10, clip_value_max=0) loss_nopt = tf.reduce_mean(error_nopt ** 2) # NOPT loss q_joint_mean = tf.reduce_mean(q_joint) elif self.config.agent == "QTRAN_alt": # -- TD Loss -- (Computed for all agents) q_count, v_joint = self.policy.Q_tran(state, hidden_state, actions, agent_mask) actions_choosen = itemgetter(*self.model_keys)(actions) actions_choosen = tf.reshape(actions_choosen, [-1, self.n_agents, 1]) q_joint_choosen = tf.reshape(tf.gather(q_count, tf.cast(actions_choosen, dtype=tf.int32), axis=-1, batch_dims=-1), [-1, self.n_agents]) q_next_count, _ = self.policy.Q_tran_target(state_next, hidden_state_next, actions_next_greedy, agent_mask) actions_next_choosen = itemgetter(*self.model_keys)(actions_next_greedy) actions_next_choosen = tf.reshape(actions_next_choosen, [-1, self.n_agents, 1]) q_joint_next_choosen = tf.reshape(tf.gather(q_next_count, tf.cast(actions_next_choosen, dtype=tf.int32), axis=-1, batch_dims=-1), [-1, self.n_agents]) y_dqn = rewards_tot + (1 - terminals_tot) * self.gamma * q_joint_next_choosen loss_td = self.mse_loss(tf.stop_gradient(y_dqn), q_joint_choosen) # TD loss # -- Opt Loss -- (Computed for all agents) q_tot_greedy = self.policy.Q_tot(q_eval_greedy_a) q_joint_greedy_hat, _ = self.policy.Q_tran(state, hidden_state, actions_greedy, agent_mask) actions_greedy_current = itemgetter(*self.model_keys)(actions_greedy) actions_greedy_current = tf.reshape(actions_greedy_current, [-1, self.n_agents, 1]) q_joint_greedy_hat_all = tf.reshape(tf.gather(q_joint_greedy_hat, tf.cast(actions_greedy_current, dtype=tf.int32), axis=-1, batch_dims=-1), [-1, self.n_agents]) error_opt = q_tot_greedy - tf.stop_gradient(q_joint_greedy_hat_all) + v_joint loss_opt = tf.reduce_mean(error_opt ** 2) # Opt loss # -- Nopt Loss -- q_eval_count = tf.reshape(itemgetter(*self.model_keys)(q_eval), [batch_size * self.n_agents, -1]) q_sums = tf.reshape(itemgetter(*self.model_keys)(q_eval_a), [-1, self.n_agents]) q_sums_repeat = tf.tile(q_sums[:, None], [1, self.n_agents, 1]) agent_mask_diag = tf.tile((1 - tf.eye(self.n_agents, dtype=tf.float32))[None], [batch_size, 1, 1]) q_sum_mask = tf.reduce_sum(q_sums_repeat * agent_mask_diag, axis=-1) q_count_for_nopt = tf.reshape(q_count, [batch_size * self.n_agents, -1]) v_joint_repeated = tf.reshape(tf.tile(v_joint, [1, self.n_agents]), [-1, 1]) error_nopt = q_eval_count + tf.reshape(q_sum_mask, [-1, 1]) - tf.stop_gradient(q_count_for_nopt) + v_joint_repeated error_nopt_min = tf.reduce_min(error_nopt, axis=-1) loss_nopt = tf.reduce_mean(error_nopt_min ** 2) # NOPT loss q_joint_mean = tf.reduce_mean(q_joint_choosen) else: raise ValueError("Mixer {} not recognised.".format(self.config.agent)) # calculate the loss function loss = loss_td + self.config.lambda_opt * loss_opt + self.config.lambda_nopt * loss_nopt gradients = tape.gradient(loss, self.policy.parameters_model) 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.parameters_model)) else: self.optimizer.apply_gradients(zip(gradients, self.policy.parameters_model)) return q_joint_mean, loss_td, loss_opt, loss_nopt, loss
# @tf.function
[docs] def learn(self, *inputs): if self.distributed_training: q_joint_mean, loss_td, loss_opt, loss_nopt, loss = self.policy.mirrored_strategy.run(self.forward_fn, args=inputs) return (self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, q_joint_mean, axis=None), self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss_td, axis=None), self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss_opt, axis=None), self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss_nopt, axis=None), self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, loss, 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=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 = tf.reshape(tf.reduce_mean(rewards[key], axis=1), [batch_size, 1]) terminals_tot = tf.reshape(tf.reduce_prod(terminals[key], axis=1), [batch_size, 1]) else: bs = batch_size rewards_tot = tf.reduce_mean(tf.stack(itemgetter(*self.agent_keys)(rewards), axis=1), axis=-1, keepdims=True) terminals_tot = tf.reduce_prod(tf.stack(itemgetter(*self.agent_keys)(terminals), axis=1), axis=1, keepdims=True) 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) q_joint_mean, loss_td, loss_opt, loss_nopt, loss = self.learn(bs, batch_size, state, obs, actions, rewards_tot, state_next, obs_next, terminals_tot, agent_mask, avail_actions, avail_actions_next, IDs) info.update({ "Q_joint": q_joint_mean.numpy(), "loss_td": loss_td.numpy(), "loss_opt": loss_opt.numpy(), "loss_nopt": loss_nopt.numpy(), "loss": loss.numpy() }) if self.iterations % self.sync_frequency == 0: self.policy.copy_target() info.update(self.callback.on_update_end(self.iterations, method="update", policy=self.policy, info=info, q_joint_mean=q_joint_mean)) return info