Adaptive neural network control of fes-induced cyclical lower leg movements

As a first step to the control of paraplegic gait by functional electrical stimulation (FES), the control of the swinging lower leg is being studied. This paper deals with a neural control system, that has been developed for this case. The control system has been tested for a model of the swinging lower leg using computer simulations. The neural controller was trained by supervised learning (SL) and by backpropagation through time (BTT). The performance of the controller with random initial weights was poor after training with BTT and fair after SL. BTT training of the neural controller with weights, which had been initialized by SL, resulted in good control. Training with BTT thus improved the performance of the controller that initially had been trained by SL. An adaptive neural control system based on BTT has been proposed and partially tested. The controller adapted relatively fast to the change of an important model parameter.