An inner-point modification of PSO for constrained optimization

In the last two decades, PSO (Particle Swarm Optimization) gained a lot of attention among the different derivative-free algorithms for global optimization. The simplicity of the implementation, compact memory usage and parallel structure represent some key features, largely appreciated. On the other hand, the absence of local information about the objective function slow down the algorithm when one or more constraints are violated, even if a penalty approach is applied. This situation becomes critical when the feasible set reduces to a small portion of the space in which the objective function needs to be investigated, and then the probability to find a feasible point by uniform sampling is small. In the present paper, a modification of the original PSO algorithm is proposed that both avoids the evaluation of the objective function outside the feasible set and preserves the parallel structure of the algorithm. Particular attention is dedicated to the parallel structure of the algorithm, in the view of its implementation on parallel architectures.