ELLIS Delft Talk: Optimistic Learning and Applications to Resource Management

This talk addresses one of the most general frameworks for modeling sequential decision problems: online learning via (non-)convex optimization. We focus on a recent and practical twist on online learning — optimism — which centers on incorporating untrusted machine-learning advice into the learning process. The goal is simple yet ambitious: boost performance when the advice proves accurate, while preserving a worst-case guarantee should it fail. We present recent results from this line of work, including optimistic learning algorithms for continuous, discrete, and stateful environments, alongside practical case studies in network optimization and control.

Sep 3, 2024