In this paper we reviewed the most important branch and bound algorithms proposed during the last decade for the capacitated vehicle routing problem with either symmetric or asymmetric cost matrix. The progress made with these algorithms with respect to those of the previous generation is considerable: the dimension of the largest instances solved has been increased from about 25 to more than 100 customers. However, the CVRP is still far from being a closed chapter in the combinatorial optimization book. In fact some Euclidean problems from the literature with 75 customers are still unsolved and, in our opinion, the size of the problems which may be actually solved in a systematic way by the present approaches is limited to few tenths of customers.
Several possible directions of research are still almost uncovered, e.g., Dantzig–Wolfe decomposition based approaches (also known as branch and price approaches), but also a more deep investigation and understanding of the capabilities of the available techniques is strongly needed. As an example, we may mention that a direct computational evaluation and comparison of the effectiveness of the algorithms presented in this paper for the symmetric case is not possible. In fact, as illustrated in Table 3, each author either considered a slightly different problem (e.g., in [20] single customers routes were not allowed, whereas Miller [40] allowed them) or solved a completely different set of instances. The only instance which has been tackled by almost all the authors we considered is the 50 customers Euclidean problem described in Christofides and Eilon [8]. However, for this instance Fisher [20] used a real-valued cost matrix with Euclidean distances, Miller [40] used an integer cost matrix with Euclidean distances rounded to the nearest integer. As to Hadjconstantinou et al. [26], they used a hybrid solution where the integer cost of each arc is defined as the Euclidean distance between its endpoints multiplied by 104 and then rounded to nearest integer. Another research issue which may lead to interesting results is represented by the adaptation to the symmetric CVRP of the exact approaches developed for the asymmetric case, and vice versa.