Movement prediction using an MLP without internal feedback

In functional electrical stimulation controllers developed based on a tracking approach, the desired movement (usually the joint angles trajectories) is used as the input of the controller. In answer to the question, how the desired movement should be individually tailored, a new method based on a multi-layer perceptron (MLP) has been developed to generate the subject-dependent trajectories of the joint angles during sit-to-stand transfer. The size of the MLP has been reduced significantly by choosing a suitable set of inputs. Outputs of the implemented MLP are the coefficients of the Fourier half amplitude cosine expansions of the joint angles. Since these coefficients describe the inherent dynamics of the system, we could avoid implementing the usual embedded feedback in the body of the MLP. In comparison with a model-based algorithm (with a maximum prediction error of 10.5%), this method predicted the movement more accurately (with a maximum error of 5.2%).