Research

Exact solvers give guarantees but scale badly; heuristics are fast but abandon them. My doctoral work put a graph neural network (GNN) inside the solver, guiding the tree search by predicting which branches matter in constrained path-planning and in disjunctive temporal networks with uncertainty (DTNU). This preserved the guarantees while cutting the explored tree substantially.

Training those models meant building the surrounding machinery as well: generators that produce problem instances on demand, and exact solvers whose proven answers supplied the ground truth to learn from. What the networks learned from solved instances transferred to unseen ones, where guided search outperformed both the unguided solver and the handcrafted heuristics written for these domains.