'main'].learning_rate = 0.00005 # called alpha in the paper
agent_params.algorithm.huber_loss_interval = 1 # k = 0 for strict quantile loss, k = 1 for Huber quantile loss
###############
# Environment #
###############
env_params = Atari()
env_params.level = SingleLevelSelection(atari_deterministic_v4)
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [
SelectedPhaseOnlyDumpMethod(RunPhase.TEST),
MaxDumpMethod()
]
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = [
'breakout', 'pong', 'space_invaders'
]
graph_manager = BasicRLGraphManager(
agent_params=agent_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params,
preset_validation_params=preset_validation_params)
agent_params.network_wrappers['actor'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'] = \
agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation')
agent_params.network_wrappers['actor'].input_embedders_parameters['measurements'].scheme = [Dense([300])]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [Dense([200])]
agent_params.network_wrappers['critic'].input_embedders_parameters['measurements'].scheme = [Dense([400])]
agent_params.network_wrappers['critic'].middleware_parameters.scheme = [Dense([300])]
agent_params.network_wrappers['critic'].input_embedders_parameters['action'].scheme = EmbedderScheme.Empty
agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter("rescale", RewardRescaleFilter(1/10.))
###############
# Environment #
###############
env_params = ControlSuiteEnvironmentParameters()
env_params.level = SingleLevelSelection(control_suite_envs)
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()]
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['cartpole:swingup', 'hopper:hop']
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=vis_params,
preset_validation_params=preset_validation_params)
agent_params.network_wrappers['main'].input_embedders_parameters['observation'] = \
InputEmbedderParameters(scheme=[Dense(200)])
agent_params.network_wrappers['main'].middleware_parameters = LSTMMiddlewareParameters(scheme=MiddlewareScheme.Empty,
number_of_lstm_cells=128)
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter('rescale', RewardRescaleFilter(1/20.))
agent_params.input_filter.add_observation_filter('observation', 'normalize', ObservationNormalizationFilter())
###############
# Environment #
###############
env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test = False
preset_validation_params.min_reward_threshold = 400
preset_validation_params.max_episodes_to_achieve_reward = 1000
preset_validation_params.num_workers = 8
preset_validation_params.reward_test_level = 'inverted_pendulum'
preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper']
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=VisualizationParameters(),
preset_validation_params=preset_validation_params)
agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise = 0
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_observation_filter('observation', 'clipping', ObservationClippingFilter(-200, 200))
agent_params.pre_network_filter = InputFilter()
agent_params.pre_network_filter.add_observation_filter('observation', 'normalize_observation',
ObservationNormalizationFilter(name='normalize_observation'))
agent_params.pre_network_filter.add_observation_filter('achieved_goal', 'normalize_achieved_goal',
ObservationNormalizationFilter(name='normalize_achieved_goal'))
agent_params.pre_network_filter.add_observation_filter('desired_goal', 'normalize_desired_goal',
ObservationNormalizationFilter(name='normalize_desired_goal'))
###############
# Environment #
###############
env_params = GymVectorEnvironment(level=SingleLevelSelection(fetch_v1))
env_params.custom_reward_threshold = -49
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['slide', 'pick_and_place', 'push', 'reach']
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=VisualizationParameters(),
preset_validation_params=preset_validation_params)
'main'].middleware_parameters = FCMiddlewareParameters()
agent_params.network_wrappers['main'].learning_rate = 0.0001
agent_params.exploration = CategoricalParameters()
###############
# Environment #
###############
env_params = Atari()
env_params.level = SingleLevelSelection(atari_deterministic_v4)
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [
SelectedPhaseOnlyDumpMethod(RunPhase.TEST),
MaxDumpMethod()
]
vis_params.dump_mp4 = False
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.trace_test_levels = ['breakout', 'pong', 'alien']
graph_manager = BasicRLGraphManager(
agent_params=agent_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params,
preset_validation_params=preset_validation_params)
