Abstract:
Bayesian Experimental Design (BED) provides a principled framework for sequential data-collection under uncertainty, and is used in a wide set of domains such as clinical trials, ecological monitoring, and hyperparameter optimization. Despite its wide applicability, BED methods remain challenging to deploy in practice due to their significant computational demands. This thesis addresses these computational bottlenecks by developing amortized policy optimization methods that reduce online costs while maintaining strong performance.

We first present a case study in multi-robot active search, where we cast the problem as a decentralized batch BED task. Our proposed algorithm—based on myopic posterior sampling—enables a team of ground robots and UAVs to coordinate search across the spectrum of no-communication to full-communication conditions. We demonstrated the robustness and efficiency of this system through extensive field tests in a forested environment.

However, despite this success, we found that we had to trade off decision-making quality to stay within real-time computational budgets. Towards addressing these computational bottlenecks, we introduce two complementary contributions to improve the scalability and robustness of learned BED policies. First, we leverage domain-permutation equivariance in BED problems to design more sample-efficient policy and Q-value architectures. Second, we propose a model-based framework for quantifying Q-value uncertainty, improving the reliability of policy optimization under model misspecification.

Together, these contributions aim to broaden the practical scope of BED by enabling real-time, high-performance decision-making in complex environments. We conclude by outlining future work on deploying these methods in real-world BED applications and extending our work on model-based Q-value uncertainty quantification.

Thesis Committee Members:
Jeff Schneider (Chair)
David Held
Aviral Kumar
Matthias Seeger (Amazon)

More Information