from rl_coach.architectures.tensorflow_components.layers import Dense
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(10000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
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 Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(10000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
schedule_params.evaluation_steps = EnvironmentEpisodes(5)
schedule_params.heatup_steps = EnvironmentSteps(50000)
#########
# Agent #
#########
agent_params = NECAgentParameters(scheme=MiddlewareScheme.RP, dense_layer=RPDense)
# If the rp_to_fc_step is inf, use only RP layers. Also if it's 0 or None, use FC layers only (same as NEC).
agent_params.rp_to_fc_step = 100000000
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 #
########
from rl_coach.agents.nec_agent import NECAgentParameters
from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps, RunPhase
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(10000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(2000)
#########
# Agent #
#########
agent_params = NECAgentParameters()
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()
env_params.level = SingleLevelSelection(atari_deterministic_v4)
env_params.random_initialization_steps = 1
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [
SelectedPhaseOnlyDumpMethod(RunPhase.TEST),
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(10)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
schedule_params.heatup_steps = EnvironmentSteps(1300)
#########
# Agent #
#########
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'