MICPSO: A method for incorporating dependencies into discrete particle swarm optimization

In this work, we present an extension to the recently developed Integer and Categorical Particle Swarm Optimization (ICPSO), which we refer to as Markovian ICPSO (MICPSO). MICPSO uses a Markov network to represent a particle's position, thus allowing each particle to incorporate information about dependencies between solution variables. In this work, we compare MICPSO to ICPSO, Integer PSO (IPSO), an Estimation of Distribution Algorithm called Markovianity-Based Optimization Algorithm (MOA), and a hillclimber on a set of benchmark vertex coloring problems. We find that MICPSO significantly outperforms all alternatives on all problems tested.