The role of sampling_weights.

[Yingdong-Hu]

Hi, what's the function of sampling_weights in trainer.yaml ? sampling_weights: [0.125, 0.125, 0.25, 0.5]

It seems that weights=1 are always used for all the episodes in function _sample_episodes_segments in dataset.py.

[eloialonso]

Hi, this is a simple way to sample recent experience more often.

In _sample_episodes_segments, we use equal weights (no effect) if weights is None or if there isn't enough episodes in the dataset.

Otherwise, the nth weight corresponds to the probability of sampling an episode in the nth chunk of the dataset in chronological order.

For instance, with weights = [0.125, 0.125, 0.25, 0.5], there is a 0.5 probability that an episode from the most recent quarter of the dataset will be sampled.

[Yingdong-Hu]

I understand, thank you!

[eloialonso]

Sorry for the misunderstanding, this was indeed a bug (reported again in #18). We removed this feature as it was not used during our experiments.