num_output_head_copies = 20 #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = EnvironmentEpisodes(2000) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(10) schedule_params.evaluation_steps = EnvironmentEpisodes(1) schedule_params.heatup_steps = EnvironmentSteps(N) #################### # DQN Agent Params # #################### agent_params = BootstrappedDQNAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00025 agent_params.memory.max_size = (MemoryGranularity.Transitions, 1000000) agent_params.algorithm.discount = 0.99 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(4) agent_params.network_wrappers['main'].heads_parameters[ 0].num_output_head_copies = num_output_head_copies agent_params.network_wrappers['main'].heads_parameters[ 0].rescale_gradient_from_head_by_factor = 1.0 / num_output_head_copies agent_params.exploration.bootstrapped_data_sharing_probability = 1.0 agent_params.exploration.architecture_num_q_heads = num_output_head_copies agent_params.exploration.epsilon_schedule = ConstantSchedule(0) agent_params.input_filter = NoInputFilter() agent_params.output_filter = NoOutputFilter() ###############
from rl_coach.agents.bootstrapped_dqn_agent import BootstrappedDQNAgentParameters from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters from rl_coach.environments.environment import SingleLevelSelection from rl_coach.environments.gym_environment import Atari, atari_deterministic_v4, atari_schedule from rl_coach.exploration_policies.ucb import UCBParameters from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager ######### # Agent # ######### agent_params = BootstrappedDQNAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00025 agent_params.exploration = UCBParameters() ############### # Environment # ############### env_params = Atari(level=SingleLevelSelection(atari_deterministic_v4)) ######## # 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=atari_schedule, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
from rl_coach.graph_managers.graph_manager import ScheduleParameters #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = EnvironmentSteps(50000000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(250000) schedule_params.evaluation_steps = EnvironmentSteps(135000) schedule_params.heatup_steps = EnvironmentSteps(50000) ######### # Agent # ######### agent_params = BootstrappedDQNAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00025 ############### # 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 # ########