Pan Xu
@pan-xu.bsky.social
Assistant Professor @DukeU | Previously @Caltech @UCLA | Machine Learning | Views are my own | he/him/his. 🌈
homepage: https://panxulab.github.io/
We propose a computationally efficient algorithm that achieves sublinear regret in online robust MDPs when the transition dynamics falls in f-divergence based uncertainty sets, and also establish matching regret lower bounds.
July 15, 2025 at 1:28 PM
We propose a computationally efficient algorithm that achieves sublinear regret in online robust MDPs when the transition dynamics falls in f-divergence based uncertainty sets, and also establish matching regret lower bounds.
We identify a unique information deficit issue inherent to this setting. We show that when the supremal visitation ratio, a quantity that measures the mismatch between the training dynamics and the deployment dynamics, is unbounded, online learning becomes exponentially hard.
July 15, 2025 at 1:28 PM
We identify a unique information deficit issue inherent to this setting. We show that when the supremal visitation ratio, a quantity that measures the mismatch between the training dynamics and the deployment dynamics, is unbounded, online learning becomes exponentially hard.