Deep Learning for Control: a non-Reinforcement Learning View

Deep learning platforms have become hugely popular due to their successes in natural language processing and image processing. Our objective is to show how deep learning platforms can be used for control problems. We do not make judgments about their performance as compared to traditional control approaches. We show that the main challenge when using deep learning platforms for learning control policies for nonlinear systems is ensuring the stability of the learning algorithm that depends on the stability of the closed loop system during the learning process. We discuss two approaches for overcoming the potential instability of the optimization algorithm, and showcase them in the context of learning a stabilizing controller for an inverted pendulum.