ray.rllib.core.learner.learner_group.LearnerGroup#

class ray.rllib.core.learner.learner_group.LearnerGroup(learner_spec: ray.rllib.core.learner.learner.LearnerSpec, max_queue_len: int = 20)[source]#

Bases: object

Coordinator of Learners.

Parameters

  • learner_spec – The specification for constructing Learners.

  • max_queue_len – The maximum number of batches to queue up if doing async_update If the queue is full itwill evict the oldest batch first.

Methods

add_module(*, module_id, module_spec)

Add a module to the Learners maintained by this LearnerGroup.

additional_update(*[, reduce_fn])

Apply additional non-gradient based updates to the Learners.

async_update(batch, *[, minibatch_size, ...])

Asnychronously do gradient based updates to the Learner(s) with batch.

get_in_queue_stats()

Returns the current stats for the input queue for this learner group.

get_state()

Get the states of the first Learners.

get_weights([module_ids])

Get the weights of the MultiAgentRLModule maintained by each Learner.

load_module_state(*[, marl_module_ckpt_dir, ...])

Load the checkpoints of the modules being trained by this LearnerGroup.

load_state(path)

Loads the state of the LearnerGroup.

remove_module(module_id)

Remove a module from the Learners maintained by this LearnerGroup.

save_state(path)

Saves the state of the LearnerGroup.

set_is_module_trainable([is_module_trainable])

Sets the function that determines whether a module is trainable.

set_state(state)

Sets the states of the Learners.

set_weights(weights)

Set the weights of the MultiAgentRLModule maintained by each Learner.

shutdown()

Shuts down the LearnerGroup.

update(batch, *[, minibatch_size, ...])

Do one or more gradient based updates to the Learner(s) based on given data.

Attributes

is_local