Deep reinforcement learning for the dynamic and uncertain vehicle routing problem

Accurate and real-time tracking for real-world urban logistics has become a popular research topic in the field of intelligent transportation. While the routing of urban logistic service is usually accomplished via complex mathematical and analytical methods. However, the nature and scope of real-world urban logistics are highly dynamic, and the existing optimization technique cannot precisely formulate the dynamic characteristics of the route. To ensure customers' demands are met, planners need to respond to these changes quickly (sometimes instantaneously). This paper proposes the formulation of a novel deep reinforcement learning framework to solve a dynamic and uncertain vehicle routing problem (DU-VRP), whose objective is to meet the uncertain servicing needs of customers in a dynamic environment. Considering uncertain information about the demands of customers in this problem, the partial observation Markov decision process is designed to frequently observe the changes in customers' demands in a real-time decision support system that consists of a deep neural network with a dynamic attention mechanism. Besides, a cutting-edge reinforcement learning algorithm is presented to control the value function of the DU-VRP for better training the routing process's dynamics and uncertainty. Computational experiments are conducted considering different data sources to obtain satisfactory solutions of the DU-VRP.