def __init__(
self,
agent_params: AgentParameters,
env_params: EnvironmentParameters,
schedule_params: ScheduleParameters,
vis_params: VisualizationParameters = VisualizationParameters(),
preset_validation_params:
PresetValidationParameters = PresetValidationParameters(),
name='simple_rl_graph'):
super().__init__(name, schedule_params, vis_params)
self.agent_params = agent_params
self.env_params = env_params
self.preset_validation_params = preset_validation_params
self.agent_params.visualization = vis_params
if self.agent_params.input_filter is None:
if env_params is not None:
self.agent_params.input_filter = env_params.default_input_filter(
)
else:
# In cases where there is no environment (e.g. batch-rl and imitation learning), there is nowhere to get
# a default filter from. So using a default no-filter.
# When there is no environment, the user is expected to define input/output filters (if required) using
# the preset.
self.agent_params.input_filter = NoInputFilter()
if self.agent_params.output_filter is None:
if env_params is not None:
self.agent_params.output_filter = env_params.default_output_filter(
)
else:
self.agent_params.output_filter = NoOutputFilter()
def __init__(
self,
agents_params: List[AgentParameters],
env_params: EnvironmentParameters,
schedule_params: ScheduleParameters,
vis_params: VisualizationParameters = VisualizationParameters(),
preset_validation_params:
PresetValidationParameters = PresetValidationParameters()):
self.sess = {agent_params.name: None for agent_params in agents_params}
self.level_managers = [] # type: List[MultiAgentLevelManager]
self.top_level_manager = None
self.environments = []
self.set_schedule_params(schedule_params)
self.visualization_parameters = vis_params
self.name = 'multi_agent_graph'
self.task_parameters = None
self._phase = self.phase = RunPhase.UNDEFINED
self.preset_validation_params = preset_validation_params
self.reset_required = False
self.num_checkpoints_to_keep = 4 # TODO: make this a parameter
# timers
self.graph_creation_time = None
self.last_checkpoint_saving_time = time.time()
# counters
self.total_steps_counters = {
RunPhase.HEATUP: TotalStepsCounter(),
RunPhase.TRAIN: TotalStepsCounter(),
RunPhase.TEST: TotalStepsCounter()
}
self.checkpoint_id = 0
self.checkpoint_saver = {
agent_params.name: None
for agent_params in agents_params
}
self.checkpoint_state_updater = None
self.graph_logger = Logger()
self.data_store = None
self.is_batch_rl = False
self.time_metric = TimeTypes.EpisodeNumber
self.env_params = env_params
self.agents_params = agents_params
self.agent_params = agents_params[0] # ...(find a better way)...
for agent_index, agent_params in enumerate(agents_params):
if len(agents_params) == 1:
agent_params.name = "agent"
else:
agent_params.name = "agent_{}".format(agent_index)
agent_params.visualization = copy.copy(vis_params)
if agent_params.input_filter is None:
agent_params.input_filter = copy.copy(
env_params.default_input_filter())
if agent_params.output_filter is None:
agent_params.output_filter = copy.copy(
env_params.default_output_filter())
def __init__(self, agent_params: AgentParameters, env_params: EnvironmentParameters,
schedule_params: ScheduleParameters,
vis_params: VisualizationParameters=VisualizationParameters(),
preset_validation_params: PresetValidationParameters = PresetValidationParameters()):
super().__init__('simple_rl_graph', schedule_params, vis_params)
self.agent_params = agent_params
self.env_params = env_params
self.preset_validation_params = preset_validation_params
self.agent_params.visualization = vis_params
if self.agent_params.input_filter is None:
self.agent_params.input_filter = env_params.default_input_filter()
if self.agent_params.output_filter is None:
self.agent_params.output_filter = env_params.default_output_filter()