Introducing User Preference Using Desirability Functions in Multi-Objective Evolutionary Optimisation of Noisy Processes

Mehnen, J.^{1, a}; Trautmann, H.^{2, b}; Tiwari, A.^{1, c}

1)

Decision Engineering Centre, University Cranfield, Cranfield, United Kingdom

2)

Lehrstuhl für Computergestützte Statistik, University of Dortmund, 44221 Dortmund, Germany

a) j.mehnen@cranfield.ac.uk; b) trautmann@statistik.uni-dortmund.de; c) a.tiwari@cranfield.ac.uk

Kurzfassung

Multi-Objective Evolutionary Algorithms (MOEAs) are generally designed to find a well spread Pareto-front approximation. Often, only a small section of this front may be of practical interest. Desirability Functions (DFs) are able to describe user preferences intuitively. Furthermore, DFs can be attached to any fitness function easily. This way, desirability functions can help in guiding MOEAs without introducing additional restrictions or changes to the algorithm. The application of noisy fitness functions is not straight forward but relevant to many real-world problems. Therefore, a variant of Harrington's one-sided desirability function using expectations is introduced which takes noise into account. A deterministic strategy as well as the NSGA-II are used in combination with DF to solve a noisy Binh problem and a noisy cost estimation problem for turning processes.

Schlüsselwörter

Desirability Functions, Noise, Noisy Objectives, User preference, Multi-objective Optimisation, Evolutionary Algorithms, Applications, CEC

Veröffentlichung

In: CEC 2007, IEEE Congress on Evolutionary Computation, 25.9.-28.9. 2007, Stamford, Singapore, Tan, K. Ch.; Xu, J.-X. (Hrsg.), ISBN 1-4244-1340-0, S. 2687-2694