AutoML-Zero and the Automated Discovery of Control Algorithms 

Abstract: I will present an unusual approach to the problem of controlling a broadly construed machine, with applications to learning and control. In this approach, we use a combination of machine learning and evolutionary computation to automatically construct computer programs that let the machine perform a desired task. Importantly, our “AutoML-Zero” method starts from scratch and proceeds without human input, only obtaining information about the task by trial-and-error through the action of the machine it controls. By combining only simple primitives, such as additions and multiplications, the process explores an enormous space of possible programs, settling in the end on one that solves the task. This search process, analogous to the evolution of the brain, results in the emergence of learning and adaptation. In particular, the process was able to make an accurate quadruped robot simulator discover how to walk while handling random breakage. In my talk, I will introduce this type of primitive symbolic search, present an overview of results, and motivate further work into this exciting research direction.

Bio: Dr. Esteban Real is a staff research scientist at Google DeepMind. He obtained a physics Ph.D. from Harvard University constructing anatomically accurate machine learning models of the vertebrate retina to learn about its function. Continuing his interest in bio-inspired computing, for the last few years he has been working on automated machine learning. His current focus is on methods for computers to automatically discover learning algorithms.