Linear splines with adaptive mesh sizes for modelling nonlinear dynamic systems

A method of identifying nonlinear dynamic models is presented which exhibits fast convergence, and adapts its memory requirements to cope with the complexity of the problem. The method modifies the CMAC algorithm by replacing fixed weights by linear splines, and may be considered as a single layer neural net. The position and number of knots (points on which the spline weights are centred) are determined adaptively in a hierarchically ordered way. The number of memory locations required depends on the degree of nonlinearity of the system being modelled. The new method is compared with CMAC on modelling a nonlinear system encountered in bioengineering (the response of muscle relaxation to a relaxant drug) and is shown to achieve comparative modelling accuracies with a reduced memory space. >