Nonlinear prediction with self-organizing maps
J. Walter, H. Riter, Klaus Schulten · 1990
The problem of predicting highly nonlinear time sequence data, where the usual approach using adaptive linear regressive models encounters difficulty, is considered. For this case, the use of an adaptive covering of the state space of the process with a set of linear regressive models, each of which is only locally used, is suggested. It is shown that such an adaptive covering, together with learning of the appropriate prediction coefficients, can be realized using Kohonen's algorithm of self-organizing maps. To illustrate the method, simulation results for a set of benchmarking problems are given