Junsung Hwang
I am an accelerated B.S. candidate in Computer Science and Applied Mathematics & Statistics at Stony Brook University, double-majoring with a minor in Mathematics and graduating in 2026. My work builds reinforcement-learning and convex-optimization pipelines for spectral graph design, learning-augmented combinatorial optimization, and reliable high-performance computing.
Across SeaWulf-scale experiments and theoretical analysis, I develop reproducible HPC workflows with stability and performance guarantees. I am especially interested in RL-driven network topology co-design, spectral graph construction, and scalable optimization methods that connect rigorous theory with deployable systems.
Academic Background
Researching reinforcement learning and convex optimization for spectral graph design with reproducible, large-scale HPC workflows.
Research Statement
My research integrates spectral graph theory, reinforcement learning, and convex optimization to construct robust network topologies. I benchmark RL controllers against effective resistance heuristics and SDP relaxations, targeting improvements in algebraic connectivity and controllability.
I build reproducible HPC pipelines on the SeaWulf cluster to orchestrate large experiment suites, enabling systematic evaluation of topology design algorithms and learning-augmented combinatorial optimization.
Current work establishes contraction-metric guarantees for CTGNN-based topology co-design, unifying theoretical stability proofs with scalable reinforcement-learning implementations.
Education
Stony Brook University
B.S. Computer Science & Applied Mathematics and Statistics (double major), Mathematics minor
Accelerated 3-year graduation, GPA 3.89/4.00, Dean's List every semester
Research Areas
- Convex optimization & reinforcement learning
- Spectral graph theory & algebraic connectivity
- Learning-augmented combinatorial optimization
- High-performance & parallel computing (MPI, OpenMP, Slurm)
- Reproducible ML systems & experiment automation
Research Focus
Bridging reinforcement learning, spectral graph theory, and high-performance computing to deliver stable, scalable network design algorithms.
Research Experiences
Spectral Optimal Graph Construction via Reinforcement Learning
Reinforcement learning for spectral graph design with provable performance benchmarks.
Contraction-based RL Topology Co-Design (CTGNN Controller)
Stability-guaranteed topology control through contraction metrics and RL.
Topology-Aware Broadcast Scheduling (BBS)
Equal-contribution work on a broadcast scheduling model tuned to given network topologies.
GraphEvo: Evolutionary Design of k-regular Graphs
Open-source evolutionary search framework for high-connectivity graphs.
Spectral Optimal Graph Construction via Reinforcement Learning
I design RL agents that edit graph topologies to maximize algebraic connectivity. The pipeline benchmarks against effective-resistance heuristics and SDP relaxations, delivering about a 12% lift in \u03bb_2 on graphs with up to 2,048 nodes while remaining computationally feasible on large clusters.
Publications & Preprints
Spectral Optimal Graph Construction via Reinforcement Learning
Junsung Hwang
Preprint, arXiv:2510.xxxxx (Oct 2025)
Sole-author preprint introducing an RL framework for spectral graph construction that outperforms effective-resistance heuristics and SDP relaxations while remaining reproducible on HPC infrastructure. Technical note and benchmarks are being developed alongside the manuscript.
A New Broadcast Model for Given Topologies
Hongbo Lu, Junsung Hwang†, Bernard Tenreiro†, Nabila Jaman Tripti, Darren Hamilton, Yuefan Deng*
Submitted to The Journal of Supercomputing (May 2025)
Equal-contribution work proposing BBS, a topology-aware broadcast scheduling model with over 30% speedup versus classical strategies across diverse network topologies on SeaWulf. († equal contribution, * corresponding author). Code release pending consortium approval.
Research Software
Tools and pipelines that accompany my work on spectral graph design, reinforcement learning, and high-performance computing.
Spectral Optimal Graph Construction via Reinforcement Learning
SeaWulf-scale reinforcement-learning pipeline that edits graph topologies to maximize algebraic connectivity, benchmarking against effective-resistance heuristics and SDP relaxations.
Preprint targeted for Oct 2025 (arXiv:2510.xxxxx)
Technologies:
GraphEvo: Evolutionary Design of k-regular Graphs
Open-source evolutionary search framework with a Master Regulatory Gene crossover, MPI-based island model, and reproducible CLI/API for high-connectivity graph discovery.
Released on GitHub with CI/CD and documentation
Technologies:
BBS: Topology-Aware Broadcast Scheduling
Equal-contribution co-authorship on BBS, a broadcast model tuned to specific network topologies that delivers over 30% speedup versus classical schedulers on SeaWulf.
Submitted to The Journal of Supercomputing (May 2025); code release pending approval