Unlocking efficiency: end-to-end optimization learning for recurrent facility operational planning

• Main Authors: Lin, Yun Hui, Yin, Xiao Feng, Tian, Qingyun
• Other Authors: School of Civil and Environmental Engineering
• Format: Article
• Language: English
• Published: 2024
• Subjects: Engineering; Learning for optimization; End-to-end optimization learning
• Online Access: https://hdl.handle.net/10356/180910
• Institution: Nanyang Technological University

This paper studies a general facility operational planning problem, which involves managing a network of facilities or infrastructures (such as road sections or tolls) to serve customers or users while considering their decentralized behaviors. The objective is to optimize the service plans for each facility, taking into account that customers aim to minimize their own costs or disutilities. This problem possesses a wide array of practical applications in operations management and transportation systems. Mathematically, it is often formalized as a bilevel programming model. Due to the inherent complexity introduced by the bilevel (sometimes, hidden bilevel) structure, the resulting model is NP-hard in general. As customer demand exhibits spatial–temporal variations in real-world scenarios, service plans often necessitate re-optimization, sometimes on a rather frequent basis, to adapt to changing demand levels. This poses computational challenges due to the complexity of solving the problem, making it difficult for companies to update service plans with high quality under tight time constraints. To address this challenge, we introduce an end-to-end optimization learning framework that combines offline optimization, machine learning techniques, and customized data generation schemes. Once the learning models are developed and trained, they can directly generate near-optimal service plans using demand information as input features, without invoking external solvers/algorithms. Through computational experiments, we demonstrate that this framework delivers outstanding performance. In most cases, it can produce solutions with optimality gaps of less than 0.11% in minimal execution times. We also provide computational insights into the role of learning models during algorithm development and their impacts on different problem classes.