In this paper, we have introduced a simple and efficient Split algorithm in O(n). The algorithm uses dominance properties and can be extended to deal with a limited number of vehicles or relaxed capacity constraints. Our computational experiments show that the new algorithm is significantly faster than the usual Bellman-based approach on VRPs of a realistic size. Positive speedups are encountered when the number of deliveries per route is greater than four. For large problems with 70,000 deliveries and few routes, a speedup factor of up to 400 is observed.
There are multiple opportunities for future research. First, one can revisit existing Split-based metaheuristics, measure their new performances, and adapt them to very large-scale CVRP instances. Several neighborhood-search, neighborhood-pruning and memories techniques [4], [13], [19] and [23] are known to successfully reduce the complexity of local searches (LS) for large problems. However, the Split algorithm remained, until now, the second most important time bottleneck, and dealing with much larger instances required improvements on both fronts. With the new O(n) algorithm, one important barrier has been cleared, and we can focus on further improving the LS.