schedule_params.steps_between_evaluation_periods = EnvironmentSteps(250000)
schedule_params.evaluation_steps = EnvironmentSteps(135000)
schedule_params.heatup_steps = EnvironmentSteps(50000)
#########
# 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'
#########
# Agent #
#########
agent_params = BCAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
agent_params.memory.max_size = (MemoryGranularity.Transitions, 1000000)
# agent_params.memory.discount = 0.99
agent_params.algorithm.discount = 0.99
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(0)
agent_params.memory.load_memory_from_file_path = 'datasets/montezuma_revenge.p'
###############
# Environment #
###############
env_params = Atari()
env_params.level = 'MontezumaRevenge-v0'
env_params.random_initialization_steps = 30
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [
SelectedPhaseOnlyDumpMethod(RunPhase.TEST),
MaxDumpMethod()
]
vis_params.dump_mp4 = False
graph_manager = BasicRLGraphManager(agent_params=agent_params,
env_params=env_params,
schedule_params=schedule_params,
vis_params=vis_params)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(50000)
#########
# Agent #
#########
agent_params = NECAgentParameters(scheme=MiddlewareScheme.Shallow)
agent_params.network_wrappers['main'].learning_rate = 0.00001
agent_params.input_filter = AtariInputFilter()
agent_params.input_filter.remove_reward_filter('clipping')
###############
# Environment #
###############
env_params = Atari(level=SingleLevelSelection(atari_deterministic_v4))
env_params.random_initialization_steps = 1
########
# Test #
########
preset_validation_params = PresetValidationParameters()
preset_validation_params.test_using_a_trace_test = False
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 = NECAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
agent_params.exploration.epsilon_schedule = LinearSchedule(0.5, 0.1, 1000)
agent_params.exploration.evaluation_epsilon = 0
agent_params.algorithm.discount = 0.99
agent_params.memory.max_size = (MemoryGranularity.Episodes, 200)
agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter('rescale',
RewardRescaleFilter(1 / 200.))
###############
# Environment #
###############
env_params = Atari()
env_params.level = 'CartPole-v0'
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.test = True
preset_validation_params.min_reward_threshold = 150
