Supporting Truck Platooning Optimization via AI

Abstract: In recent years, the concept of truck platooning has garnered significant attention within the transportation community. In this work we introduced a network design problem where multiple trucks operate within the same network. The objective is to minimize the total operational cost by leveraging platooning operations. Similar to every vehicle routing problem, we formulated this as a Mixed-Integer Linear Program (MILP), which is associated with certain complexities. The utilization of binary and integer variables contributes to the integrality complexity of MILP, resulting in a discrete and non-convex feasible region. In many cases, the computational complexity increases as the number of integer variables grows, leading to an exponential expansion of possible solutions, which often results in NP-hard problems. To tackle these challenges, we introduce a framework to address the computational complexity inherent in MILP models by harnessing the potential of AI, specifically Machine Learning (ML). To the best of the authors' knowledge, there are only a few published studies focused on optimizing truck platooning using AI. By leveraging ML, we can identify underlying patterns in specific datasets, which enable accurate predictions of target values. The generalization of our methodology is ensured by using synthetic data from various combinations of operational instances. Additionally, these instances are solved using commercial solvers such as Gurobi, Baron, and CPLEX, ensuring reliable and high-quality training data for the ML models. In our case, the computationally expensive values are the binary and integer variables associated with platooning operations. By deploying supervised machine learning algorithms, we aim to predict these complicated variables, providing an upper bound in case of near optimum solutions or direct replacement of these variables in cases of optimality guarantee solutions, thereby reducing the complexity of the MILP and enabling faster computation with commercial solvers. To evaluate the impact of our approach, we compared the computational time required to solve each instance before and after incorporating the ML prediction method. Preliminary results showed a computational time reduction of up to 20%. While these results are promising, further work is underway to improve the accuracy of predictions for more complex variables, with the expectation of even greater reductions in computational time. Our methodology shows great potential in addressing the computational complexity of large-scale MILPs, aligning with the growing needs of modern transportation systems.