ray.rllib.policy.policy.Policy.compute_gradients

Policy.compute_gradients(postprocessed_batch: SampleBatch) → Tuple[List[Tuple[numpy.array | jnp.ndarray | tf.Tensor | torch.Tensor, numpy.array | jnp.ndarray | tf.Tensor | torch.Tensor]] | List[numpy.array | jnp.ndarray | tf.Tensor | torch.Tensor], Dict[str, numpy.array | jnp.ndarray | tf.Tensor | torch.Tensor]]

Computes gradients given a batch of experiences.

Either this in combination with apply_gradients() or learn_on_batch() must be implemented by subclasses.

Parameters:

postprocessed_batch – The SampleBatch object to use for calculating gradients.

Returns:

List of gradient output values. grad_info: Extra policy-specific info values.

Return type:

grads