RI: SMALL: Robust Inference and Influence in Dynamic Environments

University Of Washington, Seattle WA

Investigators

Abstract

Many application domains, such as online content recommendation, medical trials, and logistics and operations in intelligent infrastructure require the practitioner to sequentially make decisions given little information about the environment and the optimality of selected actions.On a limited measurement budget, adaptive data collection can make the difference between measuring a phenomenon, or missing it. Recent advances in machine learning have provided rich insights into ways of using past observations to guide planning of future measurements. Most state-of-the-art approaches assume observations arise from an environment with fixed probabilistic characteristics. Yet, in practice, environments such as these tend to be the exception and not the rule. To handle this, practitioners supplement existing algorithms with ad-hoc exceptions and rule of thumb mechanisms, but relying on heuristics to account for brittle assumptions may itself be brittle and call into question the entire experiment. The objective of this project is to develop algorithms with performance guarantees for collecting data adaptively in dynamic and unpredictable environments with the goal of making robust inferences, faster and cheaper. The technical agenda of this project will advance the state-of-the-art by introducing a contextual non-stochastic pure exploration framework that extends recent advances in context-free non-stochastic best-arm identification, and exploits available contextual information when possible while simultaneously giving robust guarantees in non-stochastic environments. By taking a best-of-both-worlds approach, the project will also introduce resilience into the design of algorithms for adaptive inference and influence by building model misspecification into the framework. This will allow for simple models to be used as an approximate model for inference and influence when the model is accurate, but not mislead when it is not. The framework will also extend to hyperparameter tuning of machine learning models in the more challenging setting of contextual non-stationary or streaming-data environments (e.g., federated learning on mobile devices for ``edge computing'') thereby providing theoretical advancements with utility in several application domains. The project will conduct experiments informed by real world data in its development of a rigorous theoretical framework. Accompanying the technical agenda is an integrated research and education plan that includes cross-disciplinary undergrad and grad course development leveraging the experimental platform as well as student opportunities to engage with industry. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.