mirror of
https://github.com/gryf/coach.git
synced 2025-12-17 19:20:19 +01:00
cdb8d9e518
* tests: fix multi environment variables in configci - fix multi environment vairables in configci - removing bitflip from mujoco tests - add bitflip to gym * tests: disable mujoco_a3c_lstm + fix timeout and fix docker
62 lines
3.0 KiB
Python
62 lines
3.0 KiB
Python
from rl_coach.agents.actor_critic_agent import ActorCriticAgentParameters
|
|
from rl_coach.architectures.embedder_parameters import InputEmbedderParameters
|
|
from rl_coach.architectures.middleware_parameters import LSTMMiddlewareParameters
|
|
from rl_coach.architectures.layers import Dense
|
|
from rl_coach.base_parameters import VisualizationParameters, MiddlewareScheme, PresetValidationParameters
|
|
from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps
|
|
from rl_coach.environments.environment import SingleLevelSelection
|
|
from rl_coach.environments.gym_environment import GymVectorEnvironment, mujoco_v2
|
|
from rl_coach.filters.filter import InputFilter
|
|
from rl_coach.filters.observation.observation_normalization_filter import ObservationNormalizationFilter
|
|
from rl_coach.filters.reward.reward_rescale_filter import RewardRescaleFilter
|
|
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 #
|
|
###############
|
|
env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
|
|
|
|
########
|
|
# Test #
|
|
########
|
|
preset_validation_params = PresetValidationParameters()
|
|
preset_validation_params.test = False
|
|
preset_validation_params.min_reward_threshold = 400
|
|
preset_validation_params.max_episodes_to_achieve_reward = 1000
|
|
preset_validation_params.num_workers = 8
|
|
preset_validation_params.reward_test_level = 'inverted_pendulum'
|
|
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)
|
|
|
|
|