agent_params.algorithm.beta_entropy = 0
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 1
# How many epochs to train the network using supervised methods
agent_params.algorithm.optimization_epochs = 10
agent_params.algorithm.estimate_state_value_using_gae = True
# Distributed Coach synchronization type.
agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC
agent_params.pre_network_filter = InputFilter()
agent_params.pre_network_filter.add_observation_filter(
'observation', 'normalize_observation',
ObservationNormalizationFilter(name='normalize_observation'))
###############
# Environment #
###############
env_params = GymVectorEnvironment()
env_params.level = './environment.py:DistillerWrapperEnvironment'
vis_params = VisualizationParameters()
vis_params.dump_parameters_documentation = False
vis_params.render = True
vis_params.native_rendering = True
vis_params.dump_signals_to_csv_every_x_episodes = 1
graph_manager = BasicRLGraphManager(agent_params=agent_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params)
bottom_critic.learning_rate = 0.001
bottom_critic.batch_size = 4096
agents_params = [top_agent_params, bottom_agent_params]
###############
# Environment #
###############
time_limit = 1000
env_params = GymVectorEnvironment(
level="rl_coach.environments.mujoco.pendulum_with_goals:PendulumWithGoals")
env_params.additional_simulator_parameters = {
"time_limit": time_limit,
"random_goals_instead_of_standing_goal": False,
"polar_coordinates": polar_coordinates,
"goal_reaching_thresholds": distance_from_goal_threshold
}
env_params.frame_skip = 10
env_params.custom_reward_threshold = -time_limit + 1
vis_params = VisualizationParameters()
vis_params.native_rendering = False
graph_manager = HACGraphManager(
agents_params=agents_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params,
consecutive_steps_to_run_non_top_levels=EnvironmentSteps(40))