Final Project Submission for the Reinforcement Learning course at the École Normale Supérieure (Master MVA)!
This project explores how echolocation can be incorporated in an agent’s learning process for spatial navigation. We use the Foster model [1] as the basis of our learning algorithm. This model uses temporal difference learning and an actor-critic model to learn the optimal way to get to a specified platform. It has been proven to behave similarly to rats on path navigation tasks. We inspire our echolocation system by that used by bats for object recognition and detection, and alter the Foster algorithm in order to process echolocating sound waves at each step. Our results demonstrate that the learning process can be greatly sped up by incorporating echolocation in the learning process. This study further highlights the need for more interactive agents in reinforcement learning, especially in fields such as autonomous path navigation that aim to be applied in the real world one day.