On the use of search methodologies in system identification

Ways in which advanced search methodologies from artificial intelligence and integer programming can contribute to solve the identification problem are investigated. Several such methods are reviewed, along with techniques that are introduced to increase their efficiency. It is shown that identification exercises can effectively be based on variants of the branch-and-bound procedure and that only an easily computable admissible heuristic evaluation function is missing for one to have a true A* procedure. The major innovation introduced is that a quick exploration of the state space can be made by mapping it at equidistant model structures by adding or deleting many more than one parameter at a time. This method, combined with the use of statistical tests, gives valuable information on subsets in which the best model structure should be searched for and also saves considerable computer time. Once a smaller subset that most probably contains the best model has been identified, it can be mapped in turn at a lower scale and so on until a solution has been found. The typical performances of the branch-and-bound procedure are illustrated on both simulated and industrial data