Nonlinear system identification by means of genetic programming

The paper presents a novel nonlinear identification procedure, able to select the structure and parameters of a model according to a data - driven approach. The methodology is based on genetic programming techniques. The tree-like encryption of potential models guarantees a good spread of possible solutions within the problem search-space. During the evolutionary loop, various nonlinear models, linear in parameters are generated. To increase the convergence speed, the algorithm makes use of customized genetic operators and a local optimization procedure, based on QR decomposition. The experimental trials have proven that the approach is able to provide compact and accurate models, even when poor a priori information about the model structure is available.