TD3 (#338)

rl_coach/presets/CARLA_CIL.py

@@ -123,8 +123,8 @@ agent_params.input_filter.add_observation_filter(
 
 # no exploration is used
 agent_params.exploration = AdditiveNoiseParameters()
-agent_params.exploration.noise_percentage_schedule = ConstantSchedule(0)
-agent_params.exploration.evaluation_noise_percentage = 0
+agent_params.exploration.noise_schedule = ConstantSchedule(0)
+agent_params.exploration.evaluation_noise = 0
 
 # no playing during the training phase
 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(0)

rl_coach/presets/CartPole_Dueling_DDQN.py

@@ -53,7 +53,7 @@ env_params = GymVectorEnvironment(level='CartPole-v0')
 preset_validation_params = PresetValidationParameters()
 preset_validation_params.test = True
 preset_validation_params.min_reward_threshold = 150
-preset_validation_params.max_episodes_to_achieve_reward = 250
+preset_validation_params.max_episodes_to_achieve_reward = 300
 
 graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
                                     schedule_params=schedule_params, vis_params=VisualizationParameters(),

rl_coach/presets/Fetch_DDPG_HER_baselines.py

@@ -87,9 +87,9 @@ agent_params.memory.shared_memory = True
 agent_params.exploration = EGreedyParameters()
 agent_params.exploration.epsilon_schedule = ConstantSchedule(0.3)
 agent_params.exploration.evaluation_epsilon = 0
-# they actually take the noise_percentage_schedule to be 0.2 * max_abs_range which is 0.1 * total_range
-agent_params.exploration.continuous_exploration_policy_parameters.noise_percentage_schedule = ConstantSchedule(0.1)
-agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise_percentage = 0
+# they actually take the noise_schedule to be 0.2 * max_abs_range which is 0.1 * total_range
+agent_params.exploration.continuous_exploration_policy_parameters.noise_schedule = ConstantSchedule(0.1)
+agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise = 0
 
 agent_params.input_filter = InputFilter()
 agent_params.input_filter.add_observation_filter('observation', 'clipping', ObservationClippingFilter(-200, 200))

rl_coach/presets/Mujoco_DDPG.py

@@ -15,7 +15,7 @@ schedule_params = ScheduleParameters()
 schedule_params.improve_steps = EnvironmentSteps(2000000)
 schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(20)
 schedule_params.evaluation_steps = EnvironmentEpisodes(1)
-schedule_params.heatup_steps = EnvironmentSteps(1000)
+schedule_params.heatup_steps = EnvironmentSteps(10000)
 
 #########
 # Agent #
@@ -38,7 +38,7 @@ env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
 preset_validation_params = PresetValidationParameters()
 preset_validation_params.test = True
 preset_validation_params.min_reward_threshold = 400
-preset_validation_params.max_episodes_to_achieve_reward = 1000
+preset_validation_params.max_episodes_to_achieve_reward = 3000
 preset_validation_params.reward_test_level = 'inverted_pendulum'
 preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper']
 

rl_coach/presets/Mujoco_TD3.py

@@ -0,0 +1,49 @@
+from rl_coach.agents.td3_agent import TD3AgentParameters
+from rl_coach.architectures.layers import Dense
+from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters, EmbedderScheme
+from rl_coach.core_types import EnvironmentEpisodes, EnvironmentSteps
+from rl_coach.environments.environment import SingleLevelSelection
+from rl_coach.environments.gym_environment import GymVectorEnvironment, mujoco_v2
+from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager
+from rl_coach.graph_managers.graph_manager import ScheduleParameters
+
+####################
+# Graph Scheduling #
+####################
+
+schedule_params = ScheduleParameters()
+schedule_params.improve_steps = EnvironmentSteps(1000000)
+schedule_params.steps_between_evaluation_periods = EnvironmentSteps(5000)
+schedule_params.evaluation_steps = EnvironmentEpisodes(10)
+schedule_params.heatup_steps = EnvironmentSteps(10000)
+
+#########
+# Agent #
+#########
+agent_params = TD3AgentParameters()
+agent_params.network_wrappers['actor'].input_embedders_parameters['observation'].scheme = [Dense(400)]
+agent_params.network_wrappers['actor'].middleware_parameters.scheme = [Dense(300)]
+
+agent_params.network_wrappers['critic'].input_embedders_parameters['observation'].scheme = EmbedderScheme.Empty
+agent_params.network_wrappers['critic'].input_embedders_parameters['action'].scheme = EmbedderScheme.Empty
+agent_params.network_wrappers['critic'].middleware_parameters.scheme = [Dense(400), Dense(300)]
+
+
+###############
+# Environment #
+###############
+env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
+
+########
+# Test #
+########
+preset_validation_params = PresetValidationParameters()
+preset_validation_params.test = True
+preset_validation_params.min_reward_threshold = 500
+preset_validation_params.max_episodes_to_achieve_reward = 1100
+preset_validation_params.reward_test_level = 'hopper'
+preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper']
+
+graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
+                                    schedule_params=schedule_params, vis_params=VisualizationParameters(),
+                                    preset_validation_params=preset_validation_params)

rl_coach/presets/Starcraft_CollectMinerals_A3C.py

@@ -37,9 +37,9 @@ agent_params.network_wrappers['main'].input_embedders_parameters = {
 }
 
 agent_params.exploration = AdditiveNoiseParameters()
-agent_params.exploration.noise_percentage_schedule = ConstantSchedule(0.05)
-# agent_params.exploration.noise_percentage_schedule = LinearSchedule(0.4, 0.05, 100000)
-agent_params.exploration.evaluation_noise_percentage = 0.05
+agent_params.exploration.noise_schedule = ConstantSchedule(0.05)
+# agent_params.exploration.noise_schedule = LinearSchedule(0.4, 0.05, 100000)
+agent_params.exploration.evaluation_noise = 0.05
 
 agent_params.network_wrappers['main'].batch_size = 64
 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5