The Parameter Configuration Landscape: A Case Study on Particle Swarm Optimization
Kyle Robert Harrison, Beatrice M. Ombuki-Berman, Andries Petrus Engelbrecht · 2019
It is well known that tuning a meta-heuristic optimizer is a challenging, yet rewarding process. Despite the benefits of a properly tuned optimizer, there is very little that is understood about the actual tuning process - many automated parameter tuning methods use an assumption that parameter configurations near a promising configuration will also be promising. However, this assumption has not been verified, in general. While the field of fitness landscape analysis can provide insight into the difficulty of an optimization problem, these techniques have not yet been applied to the parameter tuning problem. This paper proposes a methodology to apply standard techniques from fitness landscape analysis to the parameter configuration landscape of an arbitrary optimizer. This allows the characterization of the parameter tuning problem for an arbitrary optimizer on an arbitrary optimization problem. The proposed methodology is then investigated for the particle swarm optimization (PSO) algorithm on 20 benchmark problems in both 10 and 30 dimensions. The results indicate that the parameter configuration landscape of the PSO algorithm is globally unimodal, yet not necessarily an easy landscape to search. Furthermore, it is found that the characteristics of the PSO parameter configuration landscape do not correlate with the characteristics of the target benchmark problems.