Selection strategies for gravitational constant G in artificial physics optimisation based on analysis of convergence properties

Abstract

The gravitational constant G is a particularly important parameter in artificial physics optimisation (APO) because it influences the algorithm's convergence. APO is a population-based heuristic whose swarm at each step can be divided into two distinct subsets: a divergent subset, and a convergent subset, the former containing all individuals exhibiting divergent behaviour, and the latter all others exhibiting convergent behaviour. How APO's population is apportioned between the divergent and convergent subsets is largely determined by the value of G. Two strategies for assigning its value were studied: a constant G, and an adaptive G. The disadvantage of the constant G case is mitigated by adaptive G by tuning the swarm's distribution between the two subsets. These strategies for selecting G were tested against several benchmark functions, and the results show that APO with an adaptive G outperforms APO with a constant G.