Multi-Agent Machine Learning

Preface

For a decade I have taught a course on adaptive control. The course focused on the classical methods of system identification, using such classic texts as Ljung [1, 2]. The course addressed traditional methods of model reference adaptive control and nonlinear adaptive control using Lyapunov techniques. However, the theory had become out of sync with current engineering practice. As such, my own research and the focus of the graduate course changed to include adaptive signal processing, and to incorporate adaptive channel equalization and echo cancellation using the least mean squares (LMS) algorithm. The course name likewise changed, from “Adaptive Control” to “Adaptive and Learning Systems.” My research was still focused on system identification and nonlinear adaptive control with application to robotics. However, by the early 2000s, I had started work with teams of robots. It was now possible to use handy robot kits and low-cost microcontroller boards to build several robots that could work together. The graduate course in adaptive and learning systems changed again; the theoretical material on nonlinear adaptive control using Lyapunov techniques was reduced, replaced with ideas from reinforcement learning. A whole new range of applications developed. The teams of robots had to learn to work together and to compete.

Today, the graduate course focuses on system identification using recursive least squares techniques, some model reference adaptive control (still using ...