Solving large-scale dynamic vehicle routing problems with stochastic requests

Dynamic vehicle routing with random requests: A literature review

“…Similarly, a ''(✓)'' in column ''DIV'' means that an implicit diversion strategy is adopted. That is, in a road network, a vehicle can be diverted to a newly revealed customer, but it cannot change its direction before it arrives at the next intersection (where there may be no customer) determined by its current route (Thomas and White III, 2004;Zhang et al, 2022a).…”

“…In most relevant papers, the decisions on acceptance or rejection are not immediately made upon request arrivals; instead, decision epochs are triggered at predefined time points or when a vehicle arrives at a customer location. In Table 2, only eight papers emphasize customer responsiveness: most of them run reoptimization procedures continuously, while Hong (2012), Zou et al (2021), andZhang et al (2022a) make immediate decisions each time a new request arrives. Unlike most of the relevant papers in which service providers make decisions explicitly, Zou et al (2021) propose several cost-sharing mechanisms which take customers' decisions into account.…”

“…As a classical model for sequential decision-making, MDP is intrinsically stochastic as it aims at maximizing the expected total reward over the entire planning horizon. Solving an MDP usually requires approximating a value function that captures the expected reward-to-go (e.g., Ulmer et al, 2018a), while Zhang et al (2022a) propose a new solution method that anticipates the future reward by knapsack-based models. In the following, we discuss the VRPDSR formulations and their solution methods in detail.…”

“…They employ a reinforcement learning method to train a parameterized decision policy online, which produces robust routes efficiently. In recent work, Zhang et al (2022a) propose potential-based lookahead policies to solve an MDP formulated for the VRPDSR. They develop novel knapsack-based approximation methods to predict the expected reward-to-go, without performing extensive online or offline simulations.…”

Reinforcement Learning Approach to Stochastic Vehicle Routing Problem With Correlated Demands

“…The application of reinforcement learning in the field of SVRP is emerging as a significant area of research. Several studies have proposed various SVRP formulations and corresponding RL-based solutions (Zhang et al, 2023a;Zhou et al, 2023;Hildebrandt et al, 2023;Jin et al, 2023). However, these formulations focus on particular VRP scenarios and do not fully address the broader scope of general SVRP, which encompasses all key sources of stochasticity in VRP, such as demand, travel costs, and time windows.…”