A Real-World Application of a Many-Objective Optimisation Complexity Reduction Process

In this paper a real-world automotive engine calibration problem has been distilled into a ten-objective many-objective optimisation problem. The objectives include dynamic measures of combustion quality as well as sensitivity quantities related to a control system actuator, which exhibits significant variation. To address the computational demands of such a high-dimensional problem, use was made of parallel computing. The objective reduction process consisted of four stages and progressively reduced objective dimensionality where evidence of local objective harmony existed. It involved the advice of the calibration engineer at various stages on objective priorities and on whether to discard clusters containing solutions of no apparent interest. This process culminated in two sub-problems, one of three and one of four conflicting objectives. From the corresponding Pareto-optimal populations (POPs), visualisation together with objective priorities was used to identify preferred solutions. A comparison of the resulting POP, preferred solution and an independently generated, manually tuned calibration was made for each of the two sub-problems. In general, the preferred solution outperformed the independent calibration.