Jingfeng Wang

Jingfeng Wang

I am currently a Postdoctoral Research Scholar at the North Carolina State University.

I am on the academic job market, seeking tenure-track positions in applied mathematics or computational science starting Fall 2027.

Optimal Transport & Operations Research

← Research overview

I am interested in using optimal transport structure to develop efficient algorithms for operations research, particularly mixed-integer optimization problems with large transport subproblems.

OTBD

GPU-accelerated optimal transport for mixed-integer optimization

OTBD accelerates repeated transport subproblem evaluations using entropy-regularized optimal transport, GPU-friendly scaling, and KL projections. Applications include capacitated p-median, hub transport, and facility location problems.

An introduction to OTBD and GPU-accelerated transport optimization (2 min 32 sec).

CPU–GPU workflow

The CPU solves the master problem, while the GPU evaluates the transport subproblem through KL projections. Dual recovery produces valid cuts; optional primal recovery provides upper bounds. Cut strengthening can improve the recovered cuts, and an exact LP solve is used when the available bounds are inconclusive.

OTBD workflow: a CPU master problem exchanges decisions, cuts, and bounds with GPU transport subproblems solved using KL projections, dual recovery, and optional primal recovery.
Overview of OTBD: the master problem runs on the CPU, and transport subproblem evaluation runs on the GPU. Select the figure to view it at full size.

Computational results

In the illustrated experiments, GPU OTBD achieves nearly two orders of magnitude faster subproblem evaluation at certain problem sizes than conventional CPU subproblem solves.

Across 50 three-index hub transport instances, GPU OTBD reduces mean subproblem evaluation time by 85.7% relative to the baseline using exact CPU LP subproblem solves. It solves 47 of the 50 instances within one hour, compared with 32 for that baseline.

Left: distributions of subproblem evaluation times on a logarithmic scale across problem sizes. Right: number of instances solved within one hour, with OT-GPU reaching 47 of 50.
Left: subproblem evaluation times across instance sizes. Right: cumulative instances solved over time. The near-two-order speedup refers to subproblem evaluation at selected sizes; the right panel shows overall solve progress.