Action-based Representation for Stochastic Optimization of Complex Real-World RVRP

Logistic planning is, in some cases, still done mostly manually with supporting software tools, mainly due to the high complexity of real-world constraints. While research of the classical VRP variants is often not directly applicable to real-world logistics problems, this work deals with the problem of logistic planning faced by a particular logistics company. The studied problem can be modeled as a static, multi-trip, single-objective RVRP with multiple depots, pickup and delivery, load splitting, heterogeneous vehicles with multiple capacities, single time windows, compartmentalized cargo groups, and various incompatibility constraints. The presented research investigates whether the currently used handmade logistic plans can be automatically improved while considering the given real-world constraints. We propose a suitable problem representation based on actions, together with two mutation operators, and compare three stochastic optimization methods. The proposed optimizers achieved an average improvement of 7.7 % on real-world data.