Reinforcement Learning: Leveraging Deep Learning for Control

Abstract

Reinforcement learning is getting a lot of attention lately. People are excited about its potential to solve complex problems in areas such as robotics and automated driving, where traditional control methods can be challenging to use. In addition to deep neural nets to represent the policy, reinforcement learning lends itself to control problems because its training incorporates repeated exploration of the environment. As such exploration is time-consuming and costly or dangerous when done with actual hardware, a simulation model is often used to represent the environment. In this talk, we provide an overview of reinforcement learning and its application to teaching a robot to walk. We discuss the differences between reinforcement learning and traditional control methods. Specific topics of reinforcement learning covered in this presentation include: • Creating environment models • Crafting effective reward functions • Deploying to embedded devices through automatic code generation for CPUs and GPUs