ABSTRACT
Vaccination uptake has become the key factor that will determine our success in containing the COVID-19 pandemic. Efficient distribution of vaccines to inoculation spots is crucial to curtailing the spread of the novel coronavirus pneumonia (COVID-19) pandemic. Normally, in a big city, a huge number of vaccines need to be transported from central depot(s) through a set of satellites to widely-scattered inoculation spots by special-purpose vehicles every day. Such a large two-echelon vehicle routing problem is computationally difficult. Moreover, the demands for vaccines evolve with the epidemic spread over time, and the actual demands are hard to determine early and exactly, which not only increases the problem difficulty but also prolongs the distribution time. Based on our practical experience of COVID-19 vaccine distribution in China, we present a hybrid machine learning and evolutionary computation method, which first uses a fuzzy deep learning model to forecast the demands for vaccines for each next day, such that we can pre-distribute the forecasted number of vaccines to the satellites in advance;after obtaining the actual demands, it uses an evolutionary algorithm (EA) to route vehicles to distribute vaccines from the satellites/depots to the inoculation spots on each day. The EA saves historical problem instances and their high-quality solutions in a knowledge base, so as to capture inherent relationship between evolving problem inputs to solutions;when solving a new problem instance on each day, the EA utilizes historical solutions that perform well on the similar instances to improve initial solution quality and hence accelerate convergence. Computational results on real-world instances of vaccine distribution demonstrate that the proposed method can produce solutions with significantly shorter distribution time compared to state-of-the-arts, and hence contribute to accelerating the achievement of herd immunity. IEEE