RESCHEDULED: Women in Transportation, Prof. Lavanya Marla

The confluence of learning and optimization holds great promise for solving dynamic, online resource allocation problems under uncertainty. In this talk, I will present brief overviews of two such problems in which tools developed independently in one field have helped solve problems in the complementary field. First, we focus on the problem of collaborative routing of aircraft and unmanned vehicles, under non-stationary spatial-temporal correlations. A long unsolved question in air and ground routing has been: “if a subset of traveling vehicles can be used for exploration to update information, how should vehicles be routed to collect information most useful to minimize costs for the entire fleet”? We expand on recent advances in multi-armed bandit algorithms to solve this problem, and generate near-optimal policies. We find this could reduce travel time and fuel burn by about 5% for the fleet. Second, we study how we can improve upon long-used regret bounds in AI/ML, which are based on omniscient information about the future. Our work builds on Brown, Smith, and Sun (2010), who present an approach for information-relaxation bounds in dynamic programs. However, the computation of these penalized information-relaxation bounds (or penalized regret bounds) for large-scale systems has hitherto been discussed by many as intractable. We present tractable methods to achieve bounds for time-space network-based problems and problems modeled as block-diagonal mixed-integer programs.

Light refreshments will be provided. If you cannot join in person, register to attend via Zoom