Regret Minimization with Performative Feedback

Watch a 19-minute conference talk from the 2022 Symposium on Foundations of Responsible Computing (FORC) where UC Berkeley researcher Tijana Zrnic explores regret minimization with performative feedback in predictive modeling. Learn how deploying predictive models affects data distribution shifts and discover novel approaches for finding optimal models while maintaining low regret. Understand the unique feedback structure that distinguishes this from traditional bandit problems, as the learner receives samples from shifted distributions rather than simple bandit feedback. Explore how regret bounds can be scaled based on distribution shift complexity rather than reward function complexity, through careful exploration techniques and confidence bound construction that relies only on shift smoothness without assuming convexity.