Evolutionary Systems Identification: New Algorithmic Concepts and Applications

Abstract

System identification denotes the data drivengeneration of mathematical models for systems; theresult of a system identification algorithm consists ina mathematical description of the behaviour of theanalysed system. Evolutionary computation is a subfieldof computational intelligence that uses conceptsinspired by natural evolution; one of the most famousevolutionary techniques is the genetic algorithm, aglobal optimisation technique using aspects inspired byevolutionary biology such as selection, recombination,mutation and inheritance. This thesis concentrates onevolutionary system identification techniques based ongenetic programming (GP), an extension of the geneticalgorithm: Mathematical expressions are produced by anevolutionary process that uses the given measurementdata. The first part of this thesis describestheoretical concepts used in this work as well as ourGP implementation for the HeuristicLab framework.Concepts for monitoring population dynamics during theexecution of the GP process are also described; we hereconcentrate on genetic diversity and geneticpropagation. The application of advanced selectionprinciples and optimization stages is also explained aswell as on-line and sliding window GP variants. Thesecond part of this thesis summarises the results ofsystem identification test series; the data sets usedhere include dynamic measurement data of mechatronicalsystems as well as classification benchmark problems.The results of these tests demonstrate the ability ofthis method to produce models of high quality fordifferent kinds of machine learning problems, and alsogive insights into population dynamic processes thatoccur during the execution of a GP process.