Designing transportation systems is an extensive process, involving constant iteration between specifying modeling assumptions and solving for system performance. However, increasing system complexity pushes classical solution paradigms to their limits, thus inhibiting engineers from understanding and designing future transportation systems. This talk explores the generalizability of alternative data-driven solution paradigms––that is, how gracefully they cope with changes to modeling assumptions. The talk considers two such approaches: deep reinforcement learning (RL) and learning-guided search. Despite superior performance of deep RL in some problems, experimental findings suggest that the methods are fragile to problem variations and thus are presently not suitable for iterative design. On the other hand, new learning-guided search methods effectively accelerate state-of-the-art solvers by up to 2-7 times. Furthermore, experiments demonstrate their generalizability across problem variations, thereby indicating promise for iterative design. Applications discussed include mixed autonomy traffic, traffic signal control, vehicle routing problems, multi-robot warehousing, and integer linear programming.