Enhancing MOPSO through the guidance of ANNs

In existing work, Artificial Neural Networks (ANNs) are often used to model objective functions for Multi-Objective Particle Swarm Optimisation (MOPSO) or MOPSO is used to aid in ANN-training. We instead use an ANN to guide the optimisation algorithm by deciding if a trial solution is worthy of full evaluation. This should be particularly helpful for computationally expensive calculations. We also introduce a level of scepticism to the result produced by the ANN, to account both for inaccuracy in the ANN and the loss of performance in a MOPSO if the reinitialisation of particles is too extreme. As a case study we used a multi-objective optimisation problem that seeks to optimise the shape of an airfoil to minimise drag and maximise lift. We evaluated several different methods for training an ANN: pre-training vs live training, continuous vs single training, and varied initial training set size. For applying the ANN's output to MOPSO we looked at various levels of scepticism and verified ANN quality before applying it. Attainment surfaces were then used to compare the performance of guided and unguided MOPSOs. Our analysis showed the performance of guided MOPSO was significantly better than unguided MOPSO. We further analysed the results to derive guidance for selecting appropriate variations for specific problems.