Why is num_steps_postierior always 0? about oyster

[katerakelly]

Hi, sorry for the late reply!

We use num_steps_posterior>0 in the sparse reward pointmass experiment, see the config here: https://github.com/katerakelly/oyster/blob/44e20fddf181d8ca3852bdf9b6927d6b8c6f48fc/configs/sparse-point-robot.json

We treated this as a hyperparameter and found that for the dense reward MuJoCo tasks, 0 worked fine. The reason is likely that those tasks don't require coordinated exploration to collect context that has task information. The sparse-reward pointmass task does require coherent exploration, and performing multiple iterations of posterior sampling facilitates that.