Automated multi-objective system identification using grammar-based genetic programming

In order to use existing identification tools effectively, a user must make critical choices a priori that ultimately determine the quality of estimated models. Furthermore, while estimated models are typically optimized for a single identification criterion, engineering applications typically impose multiple performance specifications that may contradict each other. In this contribution, we develop a system identification methodology that automatically selects parametric model structures from a wide range of dynamic system models based on measured data. The problem of inferring model structures and estimating model parameters within these structures is encapsulated in a bi-level optimization problem. The optimization problem is formulated for multiple user-specified identification objectives. Finally, the range of dynamical systems considered for the optimization problem is specified using Tree Adjoining Grammar. A solution approach based on genetic programming is developed, and its asymptotic properties and computational complexity is analysed. The empirical performance of the proposed identification techniques is studied using a simulation example.