A hybrid ranking and selection procedure to solve simulation-based discrete network design problems

Abstract

With the rapid advancements of simulation techniques, simulation-based traffic management and control have gained increasing popularity. This study specifically targets the solution algorithm for simulation-based discrete network design problems (S-DNDPs). Previous research has often tackled S-DNDPs using surrogate-based optimization approaches. However, existing methods face challenges in accuracy due to discrete decision variables and insufficiently address the stochasticity and high computational costs of traffic simulators. Consequently, there is a pressing need to develop an efficient and effective method for S-DNDPs. To this end, this study introduces a novel simulation-based optimization framework, namely the ranking and selection (R&S) procedure, to solve S-DNDPs. The R&S procedure, deemed to be a variant of the enumeration method, does not rely on (approximate) gradient information of the S-DNDP. Instead, it explores all feasible solutions, intelligently allocates computational resources, and selects the optimal one accordingly. To efficiently solve S-DNDPs, a hybrid R&S procedure is proposed by incorporating a Bayesian screening procedure in a popular R&S procedure (i.e., the optimal computing budget allocation, OCBA). In addition, a self-adaptive scheme is developed to determine the computational budget in each iteration. This study demonstrated that in comparison with OCBA, the proposed R&S procedure achieves a higher lower bound of the probability of correct selection, leading to a more efficient allocation of computational resources. To address the scalability issue, this study provides a simple yet effective extension of the hybrid R&S procedure, enabling the method to be applied to large-scale problems where computational resources may not suffice to perform exhaustive evaluations. The proposed methods, including the hybrid R&S procedure and its extension, are validated on a real-world S-DNDP. Experimental results demonstrate the superior performance of the proposed methods against benchmark methods.