Optimization using swarm algorithms is a topic that has attracted research since the introduction of the particle swarm optimization (PSO) paradigm in the 1990s. Many variants of the method aim at improving convergence while avoiding local minima.

This new supervised PSO (SuPSO) algorithm integrates the concept of swarm particles with a supervisor that dynamically groups them in different clusters to make different jobs. In that way, the contrasting tasks of exploration and exploitation are assigned to different populations of particles. The subdivision is based on a topological structure, learned by the supervisor from the fitness values. At given time intervals, a new supervisor is generated using support vector machines to assign particles to the three groups of exploitation particles, diverse particles, and support particles. Since exploitation particles tend to concentrate in some small regions, particles gradually migrate, according to some probability, and a new supervisor is constructed.

The numerical results include the analysis of 28 benchmark functions against ten other variants of PSO. The comparison is not easy, considering the role of the parameters, the diversity of the particles, and the complexity. The results show overall good behavior, and no worse results than the other considered variants.

In conclusion, this is another algorithm in the PSO bag, with more pros than cons.