Incorporating robustness into Genetic Algorithm search of stochastic simulation outputs

Abstract

This paper describes a parameter design (PD) approach that incorporates Taguchi’s robustness into the Genetic Algorithm (GA) search for optimal stochastic outputs of discrete event simulation (DES). The simulation’s stochastic nature, caused by various elements of model randomness leads to varying response averages amongst simulation runs. Ignoring such variability, when ranking solution candidates in a standard GA selection scheme, may result in search convergence to bad solutions. It is clear that adopting such solutions often results in various system design and operational difficulties. The proposed approach, therefore, aims at providing settings to model control parameters at which a certain model outcome is best in performance and is less sensitive to variations in model random (noise) factors. To this end, this paper combines Taguchi’s robust design with the flexibility of simulation–evaluation to enhance the GA selection scheme and to incorporate robustness into the GA search. Taguchi measures robustness in terms of the signal-to-noise (S/N) ratio and the quality loss function (QLF), estimated through full or fractional factorial experimental designs. The proposed approach, however, utilizes the effective GA search to replace Taguchi’s experimental design with orthogonal designs, which compensates for the various shortcomings of Taguchi’s approach. Through both robustness measures, therefore, the stochastic simulation response, estimated in terms of a mean and variance based on multiple independent simulation replicates, is transformed into a scalar GA fitness evaluation. This is expected to guide the GA selection scheme to converge to a near-optimal robust parameter design. A hypothetical job shop example is used to illustrate the application of the proposed approach.