Scalable Learn-to-Optimize Algorithms for Networks

Background

Large-scale networks can adequately capture the structure of complex systems containing thousands of nodes (e.g., cyber-physical-social components) and edges (e.g., mutual interactions). Therefore, network optimization methods have become a popular approach for addressing combinatorial optimization problems in various application domains. The majority of problems are formulated as Mixed Integer Programming (MIP) models, which become NP-hard resulting from the exponential increase in the number of decision variables and the unique structure of the feasible region for large-scale problems. To overcome the computational challenges, leveraging the valuable information during the search process to embed well-informed machine learning methods in optimization algorithms has emerged as a promising research area.

Objective

This research seeks to develop scalable Learn-To-Optimize algorithms for deterministic and stochastic formulations of MIP models concerning large-scale networks. We use a set of benchmark problems and multiple city-scale networks to demonstrate the added value of our approach compared to baseline optimization algorithms in terms of solution quality, convergence rate, and computational time.

Publications

  1. Aslani, B., Mohebbi, S., Ougthon. E. (2022). The State-Of-The-Art Survey on the Application of Optimization Methods for Cyber-physical Infrastructure Networks, Available at arXiv.
  2. Aslani, B., Mohebbi, S., (2022). “Learn to decompose multi-objective optimization models for large-scale networks”, International Transactions in Operational Research, https://doi.org/10.1111/itor.13169

Funding Source

Center for Resilient and Sustainable Communities, George Mason University (2021-2023), PI, $31,640.

GRA

Babak Aslani (PhD Candidate), Peng Ren (PhD Student)

}