Reinforcement Learning for a Routing Optimization Problem. Solving a VRP with a FedEx data set.

This thesis wants to solve a Vehicle Routing Problem through the use of a non-traditional method, Reinforcement Learning. This is complemented by the resolution of the same problem through heuristic techniques and a deep analysis of the two implementations. Firstly, the problem is solved with open-source tools provided by Google, mathematical and optimization functions. Subsequently, the same problem is solved through the development of an environment and the utilization of specific Reinforcement Learning algorithms. These generate the paths of the vehicles from the warehouse to the several customers by training an agent, which decides the actions to be taken. Lastly, an economic analysis of the two proposals is carried out concentrating especially on the new method. The research shows that the traditional method optimizes the vehicles’ routes but can work well only with small sets of non-real world data, as it faces several limitations. On the other hand, Reinforcement Learning models are more complex and can work with big sets of real world data. It must be said that this study needs further refinement to provide optimal solutions, as the ones offered are not the best ones. As a matter of fact, when trying to generalize unseen data, the model is not efficient enough. However, Reinforcement Learning remains a promising way of optimizing internal business processes, which requires additional resources and study to successfully complete its tasks.