What a prioritized draw returns.
Attributes
episodeEpisodeThe chunks, batched exactly as
SequenceReplay.sample
returns them.indiceslist of intWhere each chunk came from, to hand back to
PrioritizedSequenceReplay.update_priorities.weightsTensorImportance-sampling weights,
(batch_size,), normalised so the
largest is one. A loss corrected for the sampling bias multiplies
its per-chunk terms by these.