Home·Research·Publications·Talks·Teaching·CV dark mode

Waquar Kaleem

PhD Candidate, Industrial Engineering and Operations Research · Penn State

AI-integrated discrete optimization at scale

Many planning problems pair a decision that is discrete with a cost that is only known through data. The usual response is to call the learned model as a black box inside a search, which gives the optimizer no structure to reason with. I take the other route and represent the trained network exactly inside a mixed-integer program, so the cost becomes part of the model the solver sees rather than an oracle it queries.

Doing this honestly makes the formulation large, so the work runs on to the methods that keep it solvable. That means decomposition, penalty methods for optimizing over deep rectified linear networks, and parallel implementations that reach instance sizes a single solver call cannot. The line began with integrated location and routing, where the routing cost of a cluster of customers is learned once and then reused inside the siting decision.

Publications

Talks

← back to research