Using a cerebellar model for FES control of the upper limb

Current methods for automatic feedback control of functional electrical stimulation (FES) employ classical techniques, and require much trial-and-error calibration. Given the nonlinear and time-varying nature of the system, there is a need for more sophisticated control techniques. Artificial neural networks (ANN) have shown great promise for the control of complex nonlinear systems due to their ability to learn any arbitrarily complicated mapping. Among currently used ANN structures, the cerebellar model articulation controller (CMAC) is the one that reflects most closely the behavior of the cerebellum. The authors propose the use of a self-organizing version of CMAC (SOCMAC) that utilizes available resources more efficiently. Both CMAC and SOCMAC controllers are trained to learn the behavior of a PI controller. A model of elbow flexion/extension is used to illustrate the results