Decentralized Parallel Any Colony Optimization for Distributed Memory Systems

Ant Colony System (ACS) is a well-established variant of the Ant Colony Optimization (ACO) nature inspired metaheuristic for solving combinatorial optimization problems. We present the DMACS (Distributed Memory Ant Colony System) algorithm, which is a parallelization of ACS for distributed memory architectures. The system is decentralized, with each processor running an identical agent process which administers a part of the pheromone matrix used to record the movements of simulated ants over a graph. We evaluate a Message Passing Interface (MPI) implementation of the algorithm on the well-known Travelling Salesman Problem (TSP), running on a distributed memory cluster. The results show that the algorithm scales at least as well as previous agent-based distributed implementations of ACS, without the need to sacrifice core features of the algorithm such as local search. However, our results also demonstrate that scaling the ACS algorithm to large numbers of processes in distributed memory architectures remains a significant challenge.