-
Notifications
You must be signed in to change notification settings - Fork 93
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
feat(multi-agent): implemented VDN briefly, needed to fully debug. (#41)
- Loading branch information
1 parent
6f623f0
commit 5759300
Showing
3 changed files
with
165 additions
and
1 deletion.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,141 @@ | ||
| #!/usr/bin/env python3 | ||
| # encoding: utf-8 | ||
|
|
||
| import numpy as np | ||
| import tensorflow as tf | ||
|
|
||
| from rls.algos.base.ma_off_policy import MultiAgentOffPolicy | ||
| from rls.utils.expl_expt import ExplorationExploitationClass | ||
| from rls.utils.tf2_utils import update_target_net_weights | ||
| from rls.utils.build_networks import (ValueNetwork, | ||
| DefaultRepresentationNetwork) | ||
| from rls.utils.specs import OutputNetworkType | ||
|
|
||
|
|
||
| class VDN(MultiAgentOffPolicy): | ||
| ''' | ||
| Value-Decomposition Networks For Cooperative Multi-Agent Learning, http://arxiv.org/abs/1706.05296 | ||
| TODO: RNN, multi-step, summaries, done problem | ||
| ''' | ||
|
|
||
| def __init__(self, | ||
| envspecs, | ||
|
|
||
| lr=5.0e-4, | ||
| eps_init=1, | ||
| eps_mid=0.2, | ||
| eps_final=0.01, | ||
| init2mid_annealing_step=1000, | ||
| assign_interval=2, | ||
| network_settings={ | ||
| 'share': [128], | ||
| 'v': [128], | ||
| 'adv': [128] | ||
| }, | ||
| **kwargs): | ||
| assert not any([envspec.is_continuous for envspec in envspecs]), 'VDN only support discrete action space' | ||
| super().__init__(envspecs=envspecs, **kwargs) | ||
| self.expl_expt_mng = ExplorationExploitationClass(eps_init=eps_init, | ||
| eps_mid=eps_mid, | ||
| eps_final=eps_final, | ||
| init2mid_annealing_step=init2mid_annealing_step, | ||
| max_step=self.max_train_step) | ||
| self.assign_interval = assign_interval | ||
|
|
||
| def _create_net(name, i): return ValueNetwork( | ||
| name=name+f'_{i}', | ||
| representation_net=DefaultRepresentationNetwork( | ||
| obs_spec=self.envspecs[i].obs_spec, | ||
| name=name+f'_{i}', | ||
| vector_net_kwargs=self.vector_net_kwargs, | ||
| visual_net_kwargs=self.visual_net_kwargs, | ||
| encoder_net_kwargs=self.encoder_net_kwargs, | ||
| memory_net_kwargs=self.memory_net_kwargs), | ||
| value_net_type=OutputNetworkType.CRITIC_DUELING, | ||
| value_net_kwargs=dict(output_shape=self.envspecs[i].a_dim, network_settings=network_settings) | ||
| ) | ||
|
|
||
| self.dueling_nets = [_create_net(name='dueling_net', i=i) for i in range(self.n_agents_percopy)] | ||
| self.dueling_target_nets = [_create_net(name='dueling_target_net', i=i) for i in range(self.n_agents_percopy)] | ||
| self._target_params_update() | ||
|
|
||
| self.lr = self.init_lr(lr) | ||
| self.optimizer = self.init_optimizer(self.lr) | ||
|
|
||
| [self._worker_params_dict.update(self.dueling_nets[i]._policy_models) for i in range(self.n_agents_percopy)] | ||
| [self._all_params_dict.update(self.dueling_nets[i]._all_models) for i in range(self.n_agents_percopy)] | ||
| self._all_params_dict.update(optimizer=self.optimizer) | ||
| self._model_post_process() | ||
| self.initialize_data_buffer() | ||
|
|
||
| def choose_action(self, obs, evaluation=False): | ||
| actions = [] | ||
| for i in range(self.n_agents_percopy): | ||
| if np.random.uniform() < self.expl_expt_mng.get_esp(self.train_step, evaluation=evaluation): | ||
| actions.append(np.random.randint(0, self.envspecs[i].a_dim, self.n_copys)) | ||
| else: | ||
| a = self._get_action(obs[i], self.dueling_nets[i]) | ||
| actions.append(a.numpy()) | ||
| return actions | ||
|
|
||
| @tf.function | ||
| def _get_action(self, obs, net): | ||
| with tf.device(self.device): | ||
| q_values, _ = net(obs) | ||
| return tf.argmax(q_values, axis=-1) | ||
|
|
||
| def _target_params_update(self): | ||
| if self.global_step % self.assign_interval == 0: | ||
| for i in range(self.n_agents_percopy): | ||
| update_target_net_weights(self.dueling_target_nets[i].weights, self.dueling_nets[i].weights) | ||
|
|
||
| def learn(self, **kwargs): | ||
| self.train_step = kwargs.get('train_step') | ||
| for i in range(self.train_times_per_step): | ||
| self._learn() | ||
|
|
||
| @property | ||
| def _training_variables(self): | ||
| tv = [] | ||
| for net in self.dueling_nets: | ||
| tv += net.trainable_variables | ||
| return tv | ||
|
|
||
| @tf.function | ||
| def _train(self, BATCHs): | ||
| with tf.device(self.device): | ||
| with tf.GradientTape() as tape: | ||
| q_target_all = 0 | ||
| q_target_next_max_all = 0 | ||
| reward = 0 | ||
| for i in range(self.n_agents_percopy): | ||
| reward += BATCHs[i].reward | ||
| q = self.dueling_nets[i](BATCHs[i].obs)[0] | ||
| next_q = self.dueling_nets[i](BATCHs[i].obs_)[0] | ||
| q_target = self.dueling_target_nets[i](BATCHs[i].obs_)[0] | ||
|
|
||
| q_eval = tf.reduce_sum(tf.multiply(q, BATCHs[i].action), axis=1, keepdims=True) | ||
| q_eval_all += q_eval | ||
| next_max_action = tf.argmax(next_q, axis=1, name='next_action_int') | ||
| next_max_action_one_hot = tf.one_hot(tf.squeeze(next_max_action), self.envspecs[i].a_dim, 1., 0., dtype=tf.float32) | ||
| next_max_action_one_hot = tf.cast(next_max_action_one_hot, tf.float32) | ||
|
|
||
| q_target_next_max = tf.reduce_sum( | ||
| tf.multiply(q_target, next_max_action_one_hot), | ||
| axis=1, keepdims=True) | ||
| q_target_next_max_all += q_target_next_max | ||
|
|
||
| q_target_all = tf.stop_gradient(reward + self.gamma * q_target_next_max_all) | ||
| td_error = q_target_all - q_eval_all | ||
| q_loss = tf.reduce_mean(tf.square(td_error)) | ||
| grads = tape.gradient(q_loss, self._training_variables) | ||
| self.optimizer.apply_gradients( | ||
| zip(grads, self._training_variables) | ||
| ) | ||
| self.global_step.assign_add(1) | ||
| return dict([ | ||
| ['LOSS/loss', q_loss], | ||
| ['Statistics/q_max', tf.reduce_max(q_eval_all)], | ||
| ['Statistics/q_min', tf.reduce_min(q_eval_all)], | ||
| ['Statistics/q_mean', tf.reduce_mean(q_eval_all)] | ||
| ]) |