An orthogonal ARMA identifier with automatic order estimation for biological modeling

In this paper an ARMA identification algorithm is developed for modeling biological time series data. The algorithm is based on Gram-Schmidt orthogonalization of automatically selected basis functions from a specified function space. The selection criterion is based on recursive testing of potential benefit to the model of candidate functions. The candidate functions, AR and MA terms, are tested in a pair-wise search direction until a least-squares criterion is satisfied, thereby estimating the order. Additive noise is considered and the basic algorithm extended to improve performance in noise. The algorithm is also extended to systems with inaccessible inputs (signal modeling). Modeling of biological data from speech is included, and indicates good performance. The algorithm is derived from earlier work on nonlinear systems identification.