No-Regret Linear Bandits beyond Realizability
Liu, Chong, Yin, Ming, Wang, Yu-Xiang
–arXiv.org Artificial Intelligence
We study linear bandits when the underlying reward function is not linear. Existing work relies on a uniform misspecification parameter $\epsilon$ that measures the sup-norm error of the best linear approximation. This results in an unavoidable linear regret whenever $\epsilon > 0$. We describe a more natural model of misspecification which only requires the approximation error at each input $x$ to be proportional to the suboptimality gap at $x$. It captures the intuition that, for optimization problems, near-optimal regions should matter more and we can tolerate larger approximation errors in suboptimal regions. Quite surprisingly, we show that the classical LinUCB algorithm -- designed for the realizable case -- is automatically robust against such gap-adjusted misspecification. It achieves a near-optimal $\sqrt{T}$ regret for problems that the best-known regret is almost linear in time horizon $T$. Technically, our proof relies on a novel self-bounding argument that bounds the part of the regret due to misspecification by the regret itself.
arXiv.org Artificial Intelligence
Jul-19-2023
- Country:
- North America > United States
- California > Santa Barbara County > Santa Barbara (0.14)
- Europe > United Kingdom
- England > Cambridgeshire > Cambridge (0.04)
- North America > United States
- Genre:
- Research Report (0.82)
- Technology: