Automatic Parameter Configuration of Particle Swarm Optimization by Classification of Function Features

Abstract

Metaheuristics in stochastic local search are used in numerical optimization problems in high-dimensional spaces. A characteristic of these metaheuristics is the configuration of the parameters. These parameters are essential for the optimization behavior but depend on the objective function. In this paper we introduce a new approach to automatic parameter configuration of Particle Swarm Optimization (PSO) by classifying features of the objective function. This classification utilizes a decision tree that is trained by 32 different function features. These features result from the characteristics of the underlying function landscape and of the PSO behavior. An efficient set of parameters influences the optimization in speed and performance. In literature standard configurations are introduced for different types of metaheuristics which perform a not optimal but an adequate optimization behavior for most objective functions. PSO is an example for the parameter configuration problem [2]. The swarm behavior depends mainly on the chosen parameter and leads to solutions of different quality, i.e. bad parameter sets can lead to a disadvantageous balance between exploitation and exploration. One problem by choosing the right parameter without knowledge about the objective function is to describe the characteristics of the function which are comparable to another function.