"Incorporating Prosocial Constraints and Exploiting Problem Structure in Sequential Decision-Making" - Christine Herlihy, University of Maryland, College Park

3:30 pm - 4:30 pm
Virtual (Zoom)

Christine Herlihy, a doctoral student at the University of Maryland, College Park, will deliver a talk as part of CSRAI's Young Achievers Symposium.

Watch This Talk

"Incorporating Prosocial Constraints and Exploiting Problem Structure in Sequential Decision-Making"

Sequential decision-making tasks canonically feature an agent who must explore the environment and exploit the knowledge it gains to maximize total expected reward over some time horizon. However, when algorithms are used to make decisions and induce behaviors over time in high-stakes domains, it is often necessary to trade-off reward maximization against competing objectives, such as individual or group fairness, cooperation, or risk mitigation. Additionally, when the decisions to be made are combinatorial, careful use of the structural information which characterizes or connects our decision points may facilitate our search for efficient solutions.

In this talk, Herlihy considers a constrained resource allocation task characterized by: (1) the presence of multiple objectives; and (2) the need to exploit different types of structure contained within the problem instances in order to ensure tractability and exploit externalities. She specifically considers the restless bandit setting, where a decision-maker is tasked with determining which subset of individuals (referred to as arms) should receive a beneficial intervention at each timestep, subject to the satisfaction of a budget constraint. Each restless arm is formalized as a Markov decision process (MDP), and receipt of the intervention results in an increased probability of a favorable state transition at the next timestep, relative to lack of receipt. Her group’s core contributions include the introduction of novel algorithms to address two limitations of Whittle index-based policies, including (1) the lack of distributive fairness guarantees; and (2) the inability to exploit externalities when resources are allocated within a community.

About the Speaker

Christine Herlihy is a Ph.D. student in computer science at the University of Maryland, College Park, where she is advised by John P. Dickerson. Her research interests include sequential decision-making under uncertainty (i.e., multi-armed bandits; reinforcement learning), algorithmic fairness, knowledge representation and reasoning, and health care. During her Ph.D., she has interned at Amazon Robotics, Microsoft Research, and Google Research. She earned her M.S. from Georgia Tech and completed her undergraduate studies at Georgetown University. To learn more, you can visit her website.

About the Young Achievers Symposium

The Young Achievers Symposium highlights early career researchers in diverse fields of AI for social impact. The symposium series seeks to focus on emerging research, stimulate discussions, and initiate collaborations that can advance research in artificial intelligence for societal benefit. All events in the series are free and open to the public unless otherwise noted. Penn State students, postdoctoral scholars, and faculty with an interest in socially responsible AI applications are encouraged to attend.