Learnability for Dynamical Systems

Abstract

This paper takes a step towards defining what it means for a dynamical system, such as a robot, to be able to learn through the notion of learnability. It takes a system-theoretic view to define what learning is and establishes when a system can and can not learn. Equipped with this definition of learnability, we provide a learnability result for linear systems with quadratic costs that is then applied to two different types of mobile robots, namely a simulated two-wheel inverted pendulum robot and a real, locomoting robotic platform.