Research Overview
Current AI and machine learning paradigms largely focus on training models in silico—optimizing for isolated, static benchmarks without accounting for the complex environments they will eventually inhabit. My group takes a different path, developing principled approaches to decision-making that are designed from the ground up for the complex, dynamic, multi-agent, and uncertain environments that models will inevitably face in the real world.For an up-to-date record of our research, please check my google scholar. Our research is organized into three core areas:
Foundations for Learning in Strategic Environments
Standard machine learning intuition often breaks down in strategic settings where data is not i.i.d and instead is generated from strategic agents who may be adapting in real-time. We address this by designing simple, principled algorithmic foundations that yield robust and generalizable strategies in dynamic, multi-agent environments. To do so, we draw on tools from convex optimization, game theory, behavioral economics, and dynamical systems.
Scalable Approaches for Algorithmic Decision-Making
We bridge the gap between theoretical foundations and large-scale applications. Our work focuses on scaling up our algorithmic principles to create powerful learning algorithms for AI agents. This includes developing robust benchmarking suites and creating novel methods to incorporate prior knowledge into decision rules. This line of work ensures that our theoretical guarantees can be translated into high performing models.
Broader Impacts in Real-World Systems
We study the broader implications of automated decision-making on the systems and stakeholders they affect. Moving beyond model accuracy, our research investigates the feedback loops created by the deployment of AI models and learning algorithms in society. Through work on algorithmic collective action and strategic classification, we analyze how AI-assisted decision-making reshapes incentives and impacts the real-world environments they're deployed in.
