Desynchronization in distributed Ant Colony Optimization in HPC environment

Abstract

Metaheuristics have significant computing requirements, in particular Ant Colony Optimization (ACO) processes a population of individuals (agents/ants) roaming in a graph, leaving the pheromone trails and getting inspired by its amount perceived on the edges. If the considered problem instance is large or the time is crucial, one can try to leverage parallel, hybrid or distributed infrastructure, but the algorithm itself must be properly prepared to deal with new possibilities. We have already presented a method for efficient implementation of distributed ACO, in this paper we follow up with introducing planned desynchronization in the pheromone matrix updates in order to further increase the scalability of the proposed system. The proposed modifications allowed the algorithm to scale up to 400 computations nodes without a significant impact on results quality. Efficacy of the algorithm outperforms the standard Max–Min Ant System by 10%.