models/research/efficient-hrl/configs/train_uvf.gin

#-*-Python-*-
# Create replay_buffer
agent/CircularBuffer.buffer_size = 200000
meta/CircularBuffer.buffer_size = 200000
agent/CircularBuffer.scope = "agent"
meta/CircularBuffer.scope = "meta"

# Config train
train_uvf.environment = @create_maze_env()
train_uvf.agent_class = %AGENT_CLASS
train_uvf.meta_agent_class = %META_CLASS
train_uvf.state_preprocess_class = %STATE_PREPROCESS_CLASS
train_uvf.inverse_dynamics_class = %INVERSE_DYNAMICS_CLASS
train_uvf.replay_buffer = @agent/CircularBuffer()
train_uvf.meta_replay_buffer = @meta/CircularBuffer()
train_uvf.critic_optimizer = @critic/AdamOptimizer()
train_uvf.actor_optimizer = @actor/AdamOptimizer()
train_uvf.meta_critic_optimizer = @meta_critic/AdamOptimizer()
train_uvf.meta_actor_optimizer = @meta_actor/AdamOptimizer()
train_uvf.repr_optimizer = @repr/AdamOptimizer()
train_uvf.num_episodes_train = 25000
train_uvf.batch_size = 100
train_uvf.initial_episodes = 5
train_uvf.gamma = 0.99
train_uvf.meta_gamma = 0.99
train_uvf.reward_scale_factor = 1.0
train_uvf.target_update_period = 2
train_uvf.num_updates_per_observation = 1
train_uvf.num_collect_per_update = 1
train_uvf.num_collect_per_meta_update = 10
train_uvf.debug_summaries = False
train_uvf.log_every_n_steps = 1000
train_uvf.save_policy_every_n_steps =100000

# Config Optimizers
critic/AdamOptimizer.learning_rate = 0.001
critic/AdamOptimizer.beta1 = 0.9
critic/AdamOptimizer.beta2 = 0.999
actor/AdamOptimizer.learning_rate = 0.0001
actor/AdamOptimizer.beta1 = 0.9
actor/AdamOptimizer.beta2 = 0.999

meta_critic/AdamOptimizer.learning_rate = 0.001
meta_critic/AdamOptimizer.beta1 = 0.9
meta_critic/AdamOptimizer.beta2 = 0.999
meta_actor/AdamOptimizer.learning_rate = 0.0001
meta_actor/AdamOptimizer.beta1 = 0.9
meta_actor/AdamOptimizer.beta2 = 0.999

repr/AdamOptimizer.learning_rate = 0.0001
repr/AdamOptimizer.beta1 = 0.9
repr/AdamOptimizer.beta2 = 0.999