Online Landscape Analysis for Guiding Constraint Handling in Particle Swarm Optimisation

Many real-world optimisation problems are constrained in multiple ways. As with other metaheuristics, particle swarm optimisation (PSO) algorithms do not naturally handle constrained search spaces. When PSO is used to solve a constrained problem, then the algorithm has to be modified to incorporate an appropriate constraint handling technique. Previous studies with evolutionary algorithms have shown that the choice of the most appropriate constraint handling technique depends on the features of the problem being solved. This study investigates whether this is also the case with PSO. Results are presented to show that there is performance complementarity between different constraint handling techniques when used with a traditional global best PSO algorithm. A landscape-aware approach is then implemented that uses rules derived from offline machine learning on a training set of problem instances. The rules are used to automatically switch between constraint handling techniques during PSO search. The switching is based on landscape information collected from the particles during search and requires no additional sampling or function evaluations. Results show that the proposed approach of switching techniques performs better than using any one of the individual constraint handling techniques. It is also shown that landscape-aware switching outperforms random switching, illustrating the value of using landscape features to guide the choice of constraint handling technique for PSO.