Evolving PID-like Neurocontrollers for Nonlinear Control Problems
Stéphane Doncieux, Jean-Arcady Meyer · Control and Intelligent Systems · 2005
Abstract This article describes an empirical approach to non-linear control problems that calls upon the evolution of modular neural networks. This approach may be bootstrapped with modules that encode knowledge stemming from linear or nonlinear control theory, and it seems to be applicable to non-stationary problems as well. It has been applied here to the control of the trim and altitude of a simulated lenticular blimp that was subjected to several perturbations. The corresponding results demonstrate the superiority of the evolved networks over a hand-designed controller. They also demonstrate the capacity of evolution to exploit the intrinsic non-linearities of artificial neurons in order to generate different solutions, likely to be adapted to the context of the considered application. Key Words: evolution, neural networks, modules, lenticular blimp 1. Introduction Real world systems often exhibit non-linear and non-stationary behaviors that deeply challenge traditional control techniques of engineers like PIDs. These techniques usually rely on an analysis of a model of the system to be controlled and on an inversion around a given operating point of a linearized version of this model. As a consequence, the corresponding controllers remain effective only as long as the linearized model represents the system's actual behavior. Although several research efforts have been devoted to the extension of these mathematical approaches to non-linear and non-stationary systems [1, 2], they do not yet seem to have converged to any general conclusions, according to which a given method would clearly seem more appropriate to design the controller of a given system.