[Feature Request] Support for alpha and beta parameters' schedule in torchrl.data.PrioritizedReplayBuffer

[Mad-Chuck]

Motivation

Since in the original PER paper the parameter beta is changing its value during the training, it would be desirable to be able to perform similar experiments with torch-rl's Prioritized Experience Replay.

Solution

Being able to manually change beta or both alpha and beta by calling a method on a replay buffer or modifying its properties.

Alternatives

OpenAI baseline's implementation does it by providing an additional parameter beta during sampling (while alpha is fixed). It would be also possible to implement this with a scheduler (similar to lr schedules, like in torch.optim)

Checklist

• I have checked that there is no similar issue in the repo (required)

[vmoens]

Thanks for the suggestion.
I wonder if we should blend this in the sampler or create a separate scheduler class that handles it?
The first option would just be "give me an init and final value, and I will do the decay", whereas the second option is closer to the schedulers in pytorch's optim module (lets you do any scheduling you want).
Oc the first step would just be to do a linear annealing module like in the paper.

Opinions?

[Mad-Chuck]

Sorry for being inactive, I had been busy with my master's thesis. The more I think about this, the more I lean towards making beta just an argument in the sampling method because this parameter is only used for calculating the weights, which is done only after sampling. Since PyTorch is only a tool for machine learning, I believe that the PER buffer should provide the user with as much freedom in adapting the parameters as possible. That being said, I also believe that this way satisfies it, because the user can adapt the beta parameter at any given time, and it is the user's duty to provide scheduling, which can be done in any way the user wants.

Alpha scheduling is another story, though. However, I don't think I've ever come across a paper that discusses alpha scheduling. I'm not yet familiar with PER implementation, but I wouldn't be surprised if alpha scheduling was actually impossible or meaningless due to the fact that new priorities would undermine the sense of previous ones.