FORECASTING AND CONTROL USING ADAPTIVE CONNECTIONIST NETWORKS

We investigate the feasibility of applying connectionist networks with hidden units to forecasting and process control. We place the focus on establishing parallels with the adaptive control theory and develop a particular approach which embeds input-output pairs in a state space using delay coordinates. This approach has a theoretical basis and can, in principle, be applied to represent advance maps for nonlinear autonomous systems evolving on attractors, nonlinear control systems with state-dependent discontinuities, discrete event systems and approximate system inverses. This claim is justified in the case of discontinuous, Heaviside basis functions, by making reference to the Fourier approximation theory. A heuristic, parallel algorithm which uses error broadcasting is developed to adapt the representations. The algorithm is similar to the PERCEPTRON and ADALINE with LMS (least mean square) tuning when the hidden units are inactive. Performance is illustrated via simulation. We cancel noise generated by a deterministic system, represent a supervisor for a sequence control problem, develop a predictive controller and solve an inverseproblem. The application of connectionist networks to one-step ahead control and "input matching" requires that the high frequency gain or its inverse, be projected away from zero. This prevents the algorithm from "going to sleep". It is also necessary to constrain the parameter estimates to avoid the infinite parameter drift. The simulation examples indicate that connectionist networks and parallel computation may hold the promise of solving difficult process control and forecasting problems. © 1990.