"Nonlinear System Identification, on Algorithms Finding Initial Parameter Values for the Iterative Minimization of the Cost Function"
Jonas Sjoberg (Chalmers University, Sweden)
Computing the parameter estimate in a Maximum Likelihood system identification problem typically leads to an iterative minimization of the negative log-likelihood function. Only for models which are linear in the parameters one can be sure of avoiding local minima. A good start value for the iterative minimization decreases the risk of being trapped in a local minimum. This talks starts with a tutorial part explaining the described problem in some detail. Examples are given on earlier work on algorithms for obtaining initial estimates for linear models such as ARMAX and state space models. We then turn to nonlinear system identification. Initialization of nonlinear black-box models using incremental models is explained and motivated. Algorithms for the traditional nonlinear structures Hammerstein and Wiener models are covered and a new algorithm for initializing Wiener-Hammerstein models is presented. This new algorithm is used on the benchmark data from SYSID 2009 and the result is analyzed and explained.
Two OPTEC professors have been awarded three "Gouden Krijtjes", the yearly teaching awards given by the organization of engineering students (vtk). Prof. Lombaert was awarded the prize for the best course in civil engineering, and Prof. Diehl the prizes for the best professor and the best course in mathematical engineering (where he teaches numerical optimization). They received these awards at the yearly "proffentap" where experienced students taught them how to draft beer professionally.
