ACOTSP-MF: A memory-friendly and highly scalable ACOTSP approach

Improving Ant Colony Optimization efficiency for solving large TSP instances

“…Recently Martínez and García have introduced a similar idea of backup cities or backup lists, which contain a (constant) number of nodes which are the first of the nearest neighbors that did not fit into the respective candidate lists [33]. The backup lists are used if all nodes belonging to the candidate list of the current node have already been visited (are part of the constructed solution).…”

“…For example, modern CPUs offer rich sets of SIMD instructions for speeding up computations, but often, the implementation has to be adapted to benefit from them [47,33]. Analogously, modern GPUs can efficiently execute the ACO-based algorithms, including the ACS and MMAS, but the implementation has to consider the specifics of the GPU architecture [7,39,40].…”

“…and 32-bit floating-point numbers are used to store the pheromone values, then the pheromone matrix alone takes over 37GB of memory. Even if enough RAM is available, reads and writes to the memory matrix may still hinder the performance due to relatively slow memory access times in modern systems [33]. Peake et al [35] proposed a Restricted Pheromone Matrix in which the current pheromone values are stored only for the edges connecting a node to its nearest neighbors in the corresponding candidate list.…”

“…For example, the distances between cities can be calculated on demand, removing the necessity for storing the distance matrix. Alternatively, the precomputed distances can be stored only for the nearest neighbors of each node [33]. Overall, it is possible to reduce the memory complexity of the ACO by an order of magnitude, i.e., from O(n 2 ) to O(n).…”

“…The main motivations behind developing multiple ACO variants include adaptation to new types of problems, improvements to the convergence to high-quality solutions, and more efficient utilization of computing resources. Additionally, some work was motivated by the ever-growing computational power offered by multi-core CPUs [36,47], specialized accelerators [33,44], and general-purpose graphics processing units (GPUs) [6,12,39,40]. More computational power allowed not only to reduce the computation time but also to tackle larger problems [8,35].…”