Research

How can optimization algorithms scale with modern hardware?

GPU-accelerated optimization rethinks algorithms to use parallel computing for larger problems and faster solutions. I develop first-order methods that bring linear programming and infeasibility detection into a unified GPU framework, with convergence guarantees that connect algorithm design to reliable computation.

How can learning make constrained optimization faster and more effective?

Learning to optimize uses patterns across problems to accelerate optimization. My work spans learning to solve constrained problems directly and learning to guide classical algorithms. This includes solver-free methods for mixed-integer nonlinear optimization, and learned dual inequalities that stabilize and accelerate column generation.

How should predictive models be trained for the decisions they inform?

Decision-focused learning trains predictive models for decision quality rather than prediction accuracy alone. I developed PyEPO, a unified framework for end-to-end learning with optimization models. My work also introduces CaVE for faster decision-focused training with binary linear programs, and multi-task methods that learn across related decision problems.