Reinforcement Learning with Prototypical Representations

Learning good representations needs diverse data, and good exploration requiring diverse data needing good representations. This paper tackles this by introducing a task-agnostic pre-training stage which learns a latent space, allowing for better exploration in unseen downstream tasks.

Current solutions are task dependent.

The latent space whose elements (prototypes) are learnt through entropy based exploration using a k-NN approximation. These prototypes serve as landmarks for usnseen downstream tasks where they explore better.