####################
# DQN Agent Params #
####################
agent_params = DDQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
agent_params.network_wrappers['main'].heads_parameters = [
DuelingQHeadParameters()
]
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.exploration.epsilon_schedule = LinearSchedule(
1, 0.1, (N + 7) * 2000)
agent_params.input_filter = NoInputFilter()
agent_params.output_filter = NoOutputFilter()
###############
# Environment #
###############
env_params = GymEnvironmentParameters()
env_params.level = 'rl_coach.environments.toy_problems.exploration_chain:ExplorationChain'
env_params.additional_simulator_parameters = {
'chain_length': N,
'max_steps': N + 7
}
vis_params = VisualizationParameters()
# preset_validation_params = PresetValidationParameters()
# preset_validation_params.test = True
agent_params.network_wrappers['main'].learning_rate = 0.0001
agent_params.network_wrappers['main'].input_embedders_parameters = {
"screen": InputEmbedderParameters(input_rescaling={'image': 3.0})
}
agent_params.network_wrappers['main'].heads_parameters = [
DuelingQHeadParameters()
]
agent_params.memory.max_size = (MemoryGranularity.Transitions, 1000000)
# slave_agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(10000)
agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.1, 1000000)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(4)
agent_params.output_filter = \
OutputFilter(
action_filters=OrderedDict([
('discretization', BoxDiscretization(num_bins_per_dimension=4, force_int_bins=True))
]),
is_a_reference_filter=False
)
###############
# Environment #
###############
env_params = StarCraft2EnvironmentParameters(level='CollectMineralShards')
env_params.feature_screen_maps_to_use = [5]
env_params.feature_minimap_maps_to_use = [5]
########
# Test #
########
