I found a paper from DeepMind last year on state representation. This paper integrates state representatin learning with deep reinforcement learning. The authors try differnt loss functions related to state representation, and determine which combination of loss function gives the better performance.

This reserach is insightful since it tries to answer the question whether (or how) the state representation should be learned. One particular example is the grid cell pattern. How is the grid pattern present?


Next Post Previous Post