Automatic control is what makes engineering applications, which support our modern lifestyle, be efficient and transparent. It is due to the perpetual innovations in this field that society benefits from state of the art technology – ranging from fast and safe transportation to mobile communication via smart phones for instance – without being aware of the complicated processes at work in the background (that is why automation is commonly referred to as the invisible science). As controlling a system cannot be efficiently achieved in the absence of an adequate mathematical model, it is logical to conclude that the field of systems identification is also of crucial importance.
This work approaches the nonlinear systems identification problem, by suggesting several original genetic programming based algorithms. The proposed methods are capable of providing accurate and compact models, thus ensuring their feasibility in practical applications (such as automatic control). The genetic programming paradigm was chosen to serve as basis for this research as it is capable to handle the challenges of symbolic regression, which implies determining the models structure as well as their parameters. In realistic applications, not only is the system structure unavailable, but it is also difficult to make accurate assumptions about it, due to the profoundly nonlinear nature of most practical processes in need of a model. Due to the absence of deterministic tools (e.g. derivatives used by gradient based methods), genetic programming is more robust, namely less likely to get stuck in local optima and able to work with discontinuous objective functions.
Besides the new algorithmic enhancements and their practical use, the talk also describes the author’s theoretical endeavours, namely a schema theory variation. The latter is used to mathematically describe model feature dynamics over the generations and to provide formal assessment mechanisms for measuring the efficiency of genetic programming algorithms.