State-of-the-art and research questions: Mixed Integer Linear Programming (MILP) is widely used in various fields due to its powerful modeling capabilities and the availability of efficient solvers. It is particularly useful in network modeling because it can handle complex constraints and optimize many network-related problems. Due to the computationally intensive (NP-hard) nature of many MILP problems, solving large-scale models to optimality can quickly become overly time-consuming [1]. Evaluating a model's convergence is, therefore, critical for understanding how the solver navigates the solution space as the problem size grows and for setting appropriate termination criteria. Equally important for practical applications, tuning solver parameters based on convergence insights can greatly affect both solution quality and computation time [2]. Dolan and Moré [3] introduce performance profiles, which show the share of problems a solver can solve within a certain gap in a certain amount of time. This enables a direct comparison of gap-time curves across solvers. A survey on Branch and Bound Techniques by Huang et al. discusses solver components and notes that convergence usually slows as the search tree grows exponentially. The authors highlight the importance of analyzing node count progress alongside objective improvements [4]. Lima et al. present a column generation framework to improve convergence speed for large-scale network flow MILP [5]. Despite the existence of these solver convergence studies, little work is available to analyze how MILP convergence behaves specifically in complex network-structured problems. This research investigates the convergence behavior of a fleet assignment and network optimization model. This study offers practical insight for applying MILP methods to large-scale electric aviation planning by analyzing solver performance across increasing network sizes.
Method: The following summarizes the Fleet Assignment Network Optimization Model of Westin et al. [6]: The MILP formulation minimizes total operational cost for regional electric aviation under Public Service Obligation (PSO) constraints. Unlike traditional aviation network models that use precalculated itineraries to handle passenger flows, it contains a two-layer time-space network, one for passenger flow and one for aircraft movements. Charging times and time windows for passengers are accounted for, as well as flow conservation and capacity constraints. The problem is a form of Vehicle Routing Problem (VRP) that can be described as a one-to-one Pickup and Delivery Problem with Time Windows, a Heterogenous Fleet, Split Load, and Transshipment of the passengers modeled as a Multi-Commodity Flow Problem (MCFP). The model is implemented in Matlab and solved using Gurobi. The MILP formulation consists of a linear objective function minimizing the total cost of aircraft ownership and flight operations, subject to a set of constraints. These include a system of linear inequalities representing operational limits and equalities enforcing flow conservation across a time-expanded, two-layer network of aircraft and passenger movements. The model includes integer constraints for aircraft allocation and routing variables, as well as lower and upper bounds to restrict feasible decision variable values. The input data is based on a fixed Public Service Obligation demand matrix developed to represent hypothetical regional air service patterns in Northern Scandinavia. Each PSO route specifies the services' origin, destination, and required time frame.
The model is implemented in Matlab, where data structures such as cost vectors, constraint matrices, and variable bounds are generated. These outputs are then exported to Python and solved using Gurobi. The process uses Gurobi's callback functionality to track solver progress and log objective values over time. To evaluate scalability, the original PSO matrix with 36 routes is incrementally expanded to include networks with up to 100 routes. Due to size constraints, the model handles data in dense matrix form for smaller instances, but sparse matrices are required for larger cases, such as those with 60 or more routes. Solver convergence is analyzed by recording relevant metrics during optimization runs. These include time to first feasible solution, time to reach specific optimality gaps, total runtime, and the number of branch-and-bound nodes explored. To enable comparison across different network sizes, incumbent solutions are sampled at a 30-second interval, allowing visual analysis of both the objective value and the optimality gap over time. This approach supports an examination of how model performance evolves with increasing problem complexity.
Analysis and results: Preliminary results show a consistent but scale-sensitive convergence pattern in the MILP-based fleet assignment and network optimization model for regional electric aviation. For the baseline case of 36 routes and six airports, the solver performed efficiently, found a feasible solution in under a second, and reached a 0% optimality gap in approximately 1.75 hours. When the network was expanded to 50 and 60 routes, the solver maintained the ability to reach a 5% optimality gap within 20 to 30 minutes. However, it struggled to reduce the gap below 1% within the 4–7 hour runtime limits, indicating a significant slow-down in convergence as the model complexity increased. In the 75-route scenario, the solver's performance deteriorated. Despite running for 7 hours, only around 5,600 nodes were explored, and no feasible solution or meaningful gap reduction was achieved. This stagnation suggests that the current model configuration cannot effectively scale to large network instances without major modifications. These preliminary findings suggest that while the model remains tractable for small to mid-sized networks, its scalability is limited. Enhancing performance at larger scales will likely require the integration of advanced heuristics and/or strengthened constraints. The results provide a practical benchmark for solver limitations and offer guidance for model refinement and computational strategy in future large-scale applications.
References
[1] F. Clautiaux, and I. Ljubić, “Last fifty years of integer linear programming: A focus on recent practical advances,” European Journal of Operational Research, vol. 324, no. 3, pp. 707-731, 2025/08/01/, 2025, https://doi.org/10.1016/j.ejor.2024.11.018.
[2] S. Shashaani, D. Eckman, and S. Sanchez, “Data Farming the Parameters of Simulation-Optimization Solvers,” ACM Transactions on Modeling and Computer Simulation, vol. 34, no. 4, 2024, https://doi.org/10.1145/3680282.
[3] E. D. Dolan, and J. J. Moré, “Benchmarking optimization software with performance profiles,” Mathematical Programming, Series B, vol. 91, no. 2, pp. 201-213, 2002, https://doi.org/10.1007/s101070100263.
[4] L. Huang, X. Chen, W. Huo, J. Wang, F. Zhang, B. Bai, and L. Shi, Branch and Bound in Mixed Integer Linear Programming Problems: A Survey of Techniques and Trends, 2021, https://doi.org/10.48550/arXiv.2111.06257.
[5] V. L. de Lima, M. Iori, and F. K. Miyazawa, “Exact solution of network flow models with strong relaxations,” Mathematical Programming, vol. 197, no. 2, pp. 813-846, 2023/02/01, 2023, https://doi.org/10.1007/s10107-022-01785-9.
[6] J. Westin, L. Olsson, and P. Åhag, “Network design optimization for regional electric aviation in Northern Scandinavia,” in Proc. 2024 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM), Bangkok, Thailand, pp. 400–404, 2024, doi: 10.1109/IEEM62345.2024.10856984.