Enhancing ANN-guided MOPSO through Active Learning

Artificial Neural Networks (ANNs) have often been used to model objective functions for Multi-Objective Particle Swarm Optimisation (MOPSO); alternatively MOPSO has been used to aid in training ANNs. In previous work we instead used an ANN to guide optimisation by deciding if a trial solution was worthy of full evaluation. In this work we introduce Active Learning to the ANN-guided MOPSO. This is done by using a dynamic subset of particles from the MOPSO swarm to classify locations that are likely to be on the boundary between feasible and infeasible space. As a case study we sought to optimise the shape of an airfoil to minimise drag and maximise lift.We investigated the effect of allowing up to 20 particles from the swarm to be used for Active Learning. Our analysis showed the addition of Active Learning resulted in an increase in performance where an initial archive for training was available. However if an initial archive was not available then Active Learning performed at best equal to non-Active Learning and often worse, in some cases showing poorer performance than an unguided MOPSO.