[REQUEST] Add DroQ?

[asmith26]

I've been recently learning about Dropout-Q-Functions-for-Doubly-Efficient-Reinforcement-Learning (paper, poster) which sounds like a great RL algorithm that is "doubly (sample and computationally) efficient". It's already been used by a number of interesting projects (e.g. "A Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning, which contains code for training a simulated or real A1 quadrupedal robot to walk" and "DroQ is used in RoboPianist. It can learn to play the Super Mario Bros. theme song and The Legend of Zelda songs (Main theme and Song of storms) in 400k samples!!!:musical_keyboard::musical_keyboard:". Hence I thought it might be beneficial to add it to this library.

There are quite a few code implementations available, and I have found robopianist to be the easiest to follow (I'm no expert in this though), and it looks like SAC is heavily used too.

Not sure if DroQ is of interest to the d3rlpy community?

I'm loving d3rlpy (massive thanks!!) so would be interested in learning how to implement DroQ in d3rlpy, but I'm still quite new to the lib. Thus I very much welcome if anyone is interested in implementing this (and I could learn from any PRs etc.) or if anyone is able to provide me with any guidance regarding how I would go about creating a PR to integrate DroQ into d3rlpy (only if there is interest in integrating these of course).

Many thanks for any help!