An asymmetric basis function network for approximation of dynamical systems

This paper proposes a novel algorithm in order to approximate discrete-time dynamical systems. By using a monotone transformation of the data space, it gives asymmetric basis function (ABF) networks. Our algorithm can approximate dynamical systems using less experimental data than conventional algorithms for radial basis function (RBF) networks, and can remove numerical ill-condition problems which are bottlenecks in the conventional algorithms. An application to prediction of bifurcation phenomenon is also discussed.