Jia Wan

I am a second year PhD student in Electrical Engineering and Computer Science (EECS) at the Massachusetts Institute of Technology. I am extremely fortunate to be advised by Prof. Devavrat Shah and Prof. Martin Wainwright. I am currently working on sample-efficient reinforcement learning algorithms that exploit domain-specific problem structures.

I was previously a Rhodes Scholar at the University of Oxford where I studied statistics and worked with Prof. Jakob Foerster on coordination problems in multi-agent reinforcement learning. I was simultaneously a non-residential research fellow at the Regulation, Evaluation, and Governance Lab (RegLab) at Stanford Law School working with Prof. Daniel Ho and Prof. Jacob Goldin on uncertainty quantification for demographic disparity estimates.

Prior to Oxford, I studied mathematics and computer science at Columbia University, where I was fortunate to work with Prof. Clifford Stein and Prof. Yuri Faenza on theoretical computer science.

In my spare time I like (amateurly) composing / producing music with piano, guitar, saxophone and MIDI. I also like long-distance running, gymming, skateboarding, and performing stand-up comedy.

Email  /  CV  /  Google Scholar  /  GitHub  / 

Research

Analyzing the Sample Complexity of Model-Free Opponent Shaping

Internal Closedness and von Neumann-Morgenstern Stability in Matching Theory: Structures and Complexity

We study alternatives to stability in matching problems. We define a more general notion of stability based on the game-theoretic concept of internal stability. We focus on two families of internally stable sets of matchings: von Neumann-Morgenstern (vNM) stable and internally closed. We show that, in the marriage case, internally closed sets are an alternative to stable matchings that is as tractable as stable matchings themselves. In the roommate case, deciding if a set of matchings is internally closed or von Neumann-Morgenstern stable are both co-NP-hard.

Quantifying the Uncertainty of Demographic Disparity Estimates when Demographic Labels are Unobserved

We tackle the problem of uncertainty quantification of demographic disparity estimates when the demographic labels are unobserved. We propose a dual-bootstrap approach that explicitly accounts for imputation error, enabling inference that is valid asymptotically. Empirically, we show through simulated data that dual bootstrap produces more accurate predictions in finite samples.

Analyzing the Sample Complexity of Model-Free Opponent Shaping

We present R-FOS, a tabular version of M-FOS (model-free opponent shaping) that is more suitable for theoretical analysis. Within this discretised MDP, we adapt the Rmax algorithm as the meta-learner in the R-FOS algorithm. We derive a sample complexity bound that is exponential in the cardinality of the inner state and action space and the number of agents. Our theoretical results on scaling are supported empirically in the Matching Pennies environment.

Noisy ZSC: Breaking The Common Knowledge Assumption In Zero-Shot Coordination Games

Zero-shot coordination (ZSC) gives a proxy for human-AI coordination by preventing the agents from adopting idiosyncratic behaviors or conventions that are not interpretable by agents that were not present during training. In practice, agents might each have their private models of the tasks, which might not perfectly align with the ground truth. We propose the noisy zero-shot coordination (noisy ZSC) problem to model such coordination scenarios under misalignment. On the theoretical side, we formally introduce noisy ZSC and offer a new perspective of the noisy ZSC problem by showing that any noisy ZSC problem can be reduced to a ZSC problem with respect to a standard Dec-POMDP. On the empirical side, we propose a test bed environment called the noisy lever game to benchmark coordination under the noisy ZSC framework.

Presentations

• "Von Neumann-Morgenstern Stability and Internal Closedness in Matching Theory", Proceedings of the International Conference on Integer Programming and Combinatorial Optimization (IPCO), July 2024, Wroclaw, Poland. Link.
• "Quantifying the Uncertainty of Demographic Disparity Estimates when Demographic Labels Are Unobserved", CRIW Race, Ethnicity, and Economic Statistics for the 21st Century, National Bureau of Economic Research (NBER), March 2024. Link.

Service

• Reviewer for the 28th International Conference on Artificial Intelligence and Statistics (AISTATS 2025).
• Reviewer for the Thirteenth International Conference on Learning Representations (ICLR 2025), Singapore.
• Reviewer for the Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024), Vancouver, Canada.
• Reviewer for Aligning Reinforcement Learning Experimentalists and Theorists (ARLET) workshop at the International Conference on Machine Learning (ICML 2024), Vienna, Austria.
• Reviewer for the AI for Agent-Based Modelling (AI4ABM) workshop at the International Conference on Learning Representations (ICLR 2023), Kigali, Rwanda.

Teaching

I served as teaching assistant for the following graduate-level courses at Columbia University.

• CSOR 4231: Analysis of Algorithms
• IEOR 4573: Computational Discrete Optimization

Selected Awards

• MIT EECS Alumni Fellowship: PhD fellowship
• Rhodes Scholarship: for graduate studies in Oxford
• Albert Asher Green Memorial Prize: top academic performance across the graduating class of Columbia College, Columbia University
• Course Assistant Fellowship: excellence in teaching and academics at Columbia University Computer Science Department
• summa cum laude, Phi Beta Kappa: graduation honor
• Hong Kong Jockey Club Scholarship: merit-based full scholarship for undergraduate studies

Credit to my friend Irene for the template of this website and my profile photo.

Template from Jon Barron