ppo_pr2.gin

include 'ppo.gin'

import alf.environments.suite_socialbot

# environment config
create_environment.env_name="SocialBot-Pr2Gripper-v0"
create_environment.num_parallel_environments=60
create_environment.env_load_fn=@suite_socialbot.load

PPOLoss.entropy_regularization=0.0
PPOLoss.gamma=0.99
PPOLoss.normalize_advantages=True
PPOLoss.td_lambda=0.95
PPOLoss.td_error_loss_fn=@element_wise_squared_loss
PPOLoss.check_numerics=True

NormalProjectionNetwork.std_bias_initializer_value=-1.
actor/ActorDistributionNetwork.fc_layer_params=(100, 50, 25)
actor/ActorDistributionNetwork.activation=@alf.math.softsign

value/ValueNetwork.fc_layer_params=(100, 50, 25)

ac/AdamTF.lr=2e-4

ActorCriticAlgorithm.actor_network_ctor=@actor/ActorDistributionNetwork
ActorCriticAlgorithm.value_network_ctor=@value/ValueNetwork
Agent.optimizer=@ac/AdamTF()

# training config
TrainerConfig.mini_batch_length=1
TrainerConfig.unroll_length=100
TrainerConfig.mini_batch_size=6000
TrainerConfig.num_iterations=10000
TrainerConfig.num_updates_per_train_iter=25
TrainerConfig.eval_interval=1000
TrainerConfig.debug_summaries=True
TrainerConfig.summarize_grads_and_vars=True
TrainerConfig.summary_interval=5