Scheduling multi-configuration last-mile delivery logistics by learning from optimisation feedback and customer preferences

Last-mile delivery (LMD) logistics employ multiple delivery process configurations (e.g. depot-micro, depot-self, and depot-home) to meet the delivery time expectations of customers on a large scale. Meanwhile, scheduling delivery visits within multi-configuration LMD logistics requires solving complex, integrated, and intractable mathematical models. This paper presents a deep neuroevolution from an optimisation feedback algorithm that enables one to solve a set of decomposed configuration-based mathematical models instead. The algorithm trains a predictive model (e.g. a deep neural network) to learn to assign customers to each configuration. Then, feedback is deduced by solving a set of decomposed prescriptive models to schedule deliveries within each configuration. A single objective is minimised considering total delivery time, earliness and tardiness of deliveries, arrival deviation, and total and maximum self-pickup time. The pre-trained predictive model is compared with a surrogate prescriptive assignment model regarding computational time and optimisation feedback. The applicability of the proposed algorithm is validated by a set of stability and scalability tests based on Amazon's LMD case study. The predictive model is found to outperform the simple assignment model in 100% of the test instances. In addition, its ability to grasp contextual attributes of multiple sides in LMD logistics and generalisation is highlighted.