A high-level design process for neural-network controls through a framework of human personalities

Abstract Current learning systems in the field of feedback control deal with both unbiased nonlinearities and biased nonlinearities (where bias is measured at the origin) quite differently. Unbiased nonlinearities lend themselves to direct adaptive control methods. Biased systems, on the other hand, typically require actual learning of the bias term in order to achieve acceptable error and effort. This paper attempts to unify these approaches, and to learn to compensate for both types of nonlinearities simultaneously. To do so we utilize a graphical, quantitative theory of human personalities, which assumes that their personalities indicate how people interact with the world around them using feedback. This biologically-inspired approach allows us to develop a formal design framework for tackling this problem. Simulations with a two-link robotic manipulator demonstrate the utility of the learning design method, where gravity provides the main biased nonlinearities, while friction, centripetal, and Coriolis forces are treated as unbiased nonlinearities; our neural-network update laws learn all these robot nonlinearities at the same time. Lyapunov methods result in stability guaranties for the proposed method.