Seonho Park

Seonho Park

Postdoctoral Fellow @ Georgia Tech | Researcher @ AI4OPT

Seonho Park

Main Statement

Hello! I am Seonho Park, a Postdoctoral Fellow at the H. Milton Stewart School of Industrial and Systems Engineering at Georgia Tech. Also, I am affiliated with NSF Artificial Intelligence Research Institute for Advances in Optimization (AI4OPT). Prior to this, I obtained Ph.D. in the Industrial and Systems Engineering from the University of Florida under the supervision of Distinguished Professor Panos M. Pardalos in 2021. During Ph.D., I had been working on various topics in the intersection of optimization and machine learning including variational inference, anomaly detection, stochastic optimization, image retrieval, NLP in finance, and vehicle routing problem. Currently, I am closely working with Professor Pascal Van Hentenryck on integrating machine learning into optimization. My main focus is on speeding up optimization using End-to-end Optimization Learning, which directly estimates the optimal solution given recurring input parameters of the optimization configuration.

Working Experience

  • Postdoctoral Fellow at ISyE, Georgia Tech, AI4OPT (2021-present)
  • Research Assistant at Industrial and Systems Engineering, University of Florida (2017-2021)
  • Medical Imaging Deep Learning Research Intern at Siemens Healthineers (summer 2019)

Skills

  • Programming Languages: C++, Python, Julia
  • Optimization: Pyomo, JuMP, Ipopt, Gurobi, CPLEX, OR-Tools
  • Machine Learning: PyTorch, TensorFlow

Publications

  • Seonho Park, and Pascal Van Hentenryck. Self-supervised primal-dual learning for constrained optimization. Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 4. 2023 Main Track – Oral presentation (Paper)
  • Seonho Park, and Pascal Van Hentenryck. Self-Supervised Learning for Large-Scale Preventive Security Constrained DC Optimal Power Flow., arXiv preprint arXiv:2311.18072, 2023, (Submitted to IEEE Transactions on Power Systems) (Preprint)
  • Seonho Park, Wenbo Chen, Terrence WK Mak, and Pascal Van Hentenryck. Compact Optimization Learning for AC Optimal Power Flow., arXiv preprint arXiv:2301.08840, 2023, (To appear in IEEE Transactions on Power Systems) (Paper)
  • Seonho Park, Wenbo Chen, Dahye Han, Mathieu Tanneau, and Pascal Van Hentenryck. Confidence-Aware Graph Neural Networks for Learning Reliability Assessment Commitments., arXiv preprint arXiv:2211.15755, 2022, (To appear in IEEE Transactions on Power Systems) (Paper)
  • Wenbo Chen, Seonho Park, Mathieu Tanneau, and Pascal Van Hentenryck. Learning optimization proxies for large-scale security-constrained economic dispatch. Electric Power Systems Research 213, 2022 (Paper)
  • George Adosoglou, Seonho Park, Gianfranco Lombardo, Stefano Cagnoni, and Panos M. Pardalos. Lazy Network: A Word Embedding-Based Temporal Financial Network to Avoid Economic Shocks in Asset Pricing Models., Complexity, 2022 (Paper)
  • Seonho Park, Maciej Rysz, Kathleen M. Dipple, and Panos M. Pardalos. Homography augumented momentum constrastive learning for SAR image retrieval., arXiv preprint arXiv:2109.10329, 2021 (Paper)
  • Seonho Park, Panos M. Pardalos, Deep Data Density Estimation through Donsker-Varadhan Representation arXiv preprint arXiv:2104.06612, 2021 (Paper)
  • Seonho Park, Maciej Rysz, Kaytlin L. Fair, Panos M. Pardalos, SAR Image-based Positioning in GPS-denied Environments using Deep Cosine Similarity Neural Networks, Inverse Problems & Imaging, 2020 (Paper)
  • Seonho Park, George Adosoglou, Panos M. Pardalos, Interpreting Rate-Distortion of Variational Autoencoder and Using Model Uncertainty for Anomaly Detection, Annals of Mathematics and Artificial Intelligence, 1-18, 2020 (Paper)
  • Seonho Park, Seung Hyun Jeong, Panos M. Pardalos, Combining Stochastic Adaptive Cubic Regularization with Negative Curvature for Nonconvex Optimization, Journal of optimization theory and applications, 184, pp. 953–971, 2020 (Paper)
  • Seonho Park, Seung Hyun Jeong, Gil Ho Yoon, Albert A. Groenwold, Dong-Hoon Choi, A globally convergent sequential convex programming using an enhanced two-point diagonal quadratic approximation for structural optimization, Structural and Multidisciplinary Optimization 50 (5), pp.739-753, 2014 (Paper)

Research Highlights

1. Interplay between MILP and graph neural network

alt txt
The left panel shows the ground truth optimal commitment (binary variable represented as existence of a circle) and dispatch (continuous variable represented as a radius of a circle) given 24 hour operations with the 1 hour granularity in the French RTE system. The right panel shows the estimated optimal solution using the graph neural network based approach, which is quite similar to the ground truth. More explanation can be found Here.

2. Self-supervised training framework to find out the optimal solution estimates to the constrained optimization

When it comes to training the deep neural network based mapping to output the optimal solution estimates to the constrained optimization, the supervised learning schemes exploit the pairs of the input parameter and the corresponding optimal solution. In this work, we proposed Primal Dual Learning (PDL) to train the model without relying on the ground truth, instead it follows an optimization algorithm’s trajectories to converge to the optimal solutions. PDL shows the SOTA performance in the set of experiments for estimating the optimal solutions. This work is featured at AAAI-23. The more details can be found Here.