Parameter evolution for a particle swarm optimization algorithm

Setting appropriate parameters of an evolutionary algorithm (EA) is challenging in real world applications. On one hand, the characteristics of a real world problem are usually unknown. On the other hand, in different running stages of an EA, the best parameters may be different. Thus adaptively tuning algorithm parameters online is preferred. In this paper, we propose to use an estimation of distribution algorithm (EDA) to do this for a particle swarm optimization (PSO) algorithm. The major characteristic of our approach is that there are two evolving processes simultaneously: one for tackling the original problem, and the other for optimizing PSO parameters. For the former evolving process, a set of particles are maintained; while for the later, a probability distribution model of the PSO parameters is maintained throughout the run. In the reproduction procedure, the PSO parameters are firstly sampled from the model, and then new particles are generated by the PSO operator. The feedback from the newly generated particles is used to evaluate the PSO parameters and thus to update the probability model. The new approach is applied to a set of test instances and the preliminary results are promising.