On the Computational Tractability of Robust Bandits
Researchers give a polynomial-time robust bandit learner with near-square-root regret and show nearby variants are NP-hard.
The paper studies computational tractability of robust (formerly imprecise) bandits for unrealizable learning, where prior work gave a Θ(√T) regret learner for a large class without computational guarantees. The authors identify a special case with a polynomial-time learner achieving Õ(√T) regret, and show that several small generalizations are NP-hard, placing the case at the tractability boundary. They note prior arguments that efficient learners for unrealizable problems matter for AI alignment, but the contribution is a bandit complexity result rather than an alignment method.
- A special case of robust bandits admits a polynomial-time learner.
- That learner achieves tilde-O of square-root-T regret.
- Several small generalizations of the special case are NP-hard.
- Authors frame efficient unrealizable learning as relevant to AI alignment.
From the original article
1 min read at arxiv.orgLearning when the environment does not belong to the learner's hypothesis class is typically handled using agnostic learning guarantees. However, for anything beyond supervised learning, agnostic guarantees are difficult to come by.
By Vanessa Kosoy · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.