ray.rllib.core.learner.learner.LearnerHyperparameters#

class ray.rllib.core.learner.learner.LearnerHyperparameters(learning_rate: Optional[Union[float, List[List[Union[int, float]]]]] = None, grad_clip: Optional[float] = None, grad_clip_by: Optional[str] = None, seed: Optional[int] = None, _per_module_overrides: Optional[Dict[str, ray.rllib.core.learner.learner.LearnerHyperparameters]] = None)[source]#

Bases: object

Hyperparameters for a Learner, derived from a subset of AlgorithmConfig values.

Instances of this class should only be created via calling get_learner_hyperparameters() on a frozen AlgorithmConfig object and should always considered read-only.

When creating a new Learner, you should also define a new sub-class of this class and make sure the respective AlgorithmConfig sub-class has a proper implementation of the get_learner_hyperparameters method.

Validation of the values of these hyperparameters should be done by the respective AlgorithmConfig class.

For configuring different learning behaviors for different (single-agent) RLModules within the Learner, RLlib uses the _per_module_overrides property (dict), mapping ModuleID to a overridden version of self, in which the module-specific override settings are applied.

Methods

get_hps_for_module(module_id)

Returns a LearnerHyperparameter instance, given a module_id.

Attributes

grad_clip

grad_clip_by

learning_rate

seed