Python ActorCriticAgentParameters示例

示例#1

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 = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(20)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
schedule_params.heatup_steps = EnvironmentSteps(0)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.005
agent_params.network_wrappers['main'].learning_rate = 0.00002
agent_params.network_wrappers['main'].input_embedders_parameters['observation'] = \
    InputEmbedderParameters(scheme=[Dense(200)])
agent_params.network_wrappers['main'].middleware_parameters = LSTMMiddlewareParameters(scheme=MiddlewareScheme.Empty,
                                                                                       number_of_lstm_cells=128)

agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter('rescale', RewardRescaleFilter(1/20.))
agent_params.input_filter.add_observation_filter('observation', 'normalize', ObservationNormalizationFilter())

###############
# Environment #

示例#2

from rl_coach.graph_managers.graph_manager import ScheduleParameters
from rl_coach.schedules import LinearSchedule

####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(int(5e5))
schedule_params.steps_between_evaluation_periods = EnvironmentSteps(50000)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(150000)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()

agent_params.algorithm.policy_gradient_rescaler = PolicyGradientRescaler.GAE
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter("rescale", RewardRescaleFilter(1 / 10000.0))
agent_params.algorithm.num_steps_between_gradient_updates = 30
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 0.99
agent_params.algorithm.beta_entropy = 0.01

agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.05
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(2048)

示例#3

from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps, RunPhase
from rl_coach.exploration_policies.categorical import CategoricalParameters

####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(10000)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()

agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.05

agent_params.network_wrappers['main'].learning_rate = 0.0001
agent_params.network_wrappers['main'].middleware_parameters = LSTMMiddlewareParameters(scheme=MiddlewareScheme.Medium,
                                                                                       number_of_lstm_cells=256)
agent_params.exploration = CategoricalParameters()

###############
# Environment #
###############
env_params = Atari()
env_params.level = SingleLevelSelection(atari_deterministic_v4)

示例#4

from rl_coach.environments.doom_environment import DoomEnvironmentParameters

####################
# 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(0)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()
agent_params.algorithm.policy_gradient_rescaler = PolicyGradientRescaler.GAE
agent_params.network_wrappers['main'].learning_rate = 0.0001
agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter('rescale',
                                            RewardRescaleFilter(1 / 100.))
agent_params.algorithm.num_steps_between_gradient_updates = 30
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.gae_lambda = 1.0
agent_params.algorithm.beta_entropy = 0.01
agent_params.network_wrappers['main'].clip_gradients = 40.
agent_params.exploration = CategoricalParameters()

###############
# Environment #
###############

示例#5

文件： Starcraft_CollectMinerals_A3C.py 项目： wwxFromTju/coach

from rl_coach.graph_managers.graph_manager import ScheduleParameters
from rl_coach.schedules import ConstantSchedule

####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(50)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
schedule_params.heatup_steps = EnvironmentSteps(0)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()

agent_params.algorithm.policy_gradient_rescaler = PolicyGradientRescaler.GAE
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.gae_lambda = 0.96
agent_params.algorithm.beta_entropy = 0

agent_params.network_wrappers['main'].clip_gradients = 10.0
agent_params.network_wrappers['main'].learning_rate = 0.00001
# agent_params.network_wrappers['main'].batch_size = 20
agent_params.network_wrappers['main'].input_embedders_parameters = {
    "screen": InputEmbedderParameters(input_rescaling={'image': 3.0})
}

agent_params.exploration = AdditiveNoiseParameters()

示例#6

文件： Atari_A3C.py 项目： aryanp35/RL_Project

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 = TrainingSteps(10000000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(100)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(0)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()

agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.05

agent_params.network_wrappers[
    'main'].middleware_parameters = FCMiddlewareParameters()
agent_params.network_wrappers['main'].learning_rate = 0.0001

###############
# Environment #
###############
env_params = Atari(level=SingleLevelSelection(atari_deterministic_v4))

########

示例#7

文件： Mujoco_A3C.py 项目： danialkamran/coach

from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps, RunPhase
from rl_coach.filters.reward.reward_rescale_filter import RewardRescaleFilter

####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(20000000)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(20)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
schedule_params.heatup_steps = EnvironmentSteps(0)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 10000000
agent_params.algorithm.beta_entropy = 0.0001
agent_params.network_wrappers['main'].learning_rate = 0.00001

agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter('rescale',
                                            RewardRescaleFilter(1 / 20.))
agent_params.input_filter.add_observation_filter(
    'observation', 'normalize', ObservationNormalizationFilter())

agent_params.exploration = ContinuousEntropyParameters()

###############
# Environment #

示例#8

文件： preset-autoscale-a3c.py 项目： FNDaily/amazon-sagemaker-examples

from rl_coach.exploration_policies.e_greedy import EGreedyParameters
from rl_coach.schedules import LinearSchedule

####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = TrainingSteps(int(5e5))
schedule_params.steps_between_evaluation_periods = EnvironmentSteps(50000)
schedule_params.evaluation_steps = EnvironmentEpisodes(3)
schedule_params.heatup_steps = EnvironmentSteps(150000)

#########
# Agent #
#########
agent_params = ActorCriticAgentParameters()

agent_params.algorithm.policy_gradient_rescaler = PolicyGradientRescaler.GAE
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter('rescale', RewardRescaleFilter(1/10000.))
agent_params.algorithm.num_steps_between_gradient_updates = 30
agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = 0.99
agent_params.algorithm.beta_entropy = 0.01

agent_params.algorithm.apply_gradients_every_x_episodes = 1
agent_params.algorithm.num_steps_between_gradient_updates = 20
agent_params.algorithm.beta_entropy = 0.05
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(2048)