# Agent # ######### agent_params = QuantileRegressionDQNAgentParameters() agent_params.network_wrappers[ '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,
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)