First- and Second-Order Bounds for Adversarial Linear Contextual Bandits
| Authors |
|
|---|---|
| Publication date | 2023 |
| Host editors |
|
| Book title | 37th Conference on Neural Information Processing Systems (NeurIPS 2023) |
| Book subtitle | 10-16 December 2023, New Orleans, Louisana, USA |
| ISBN (electronic) |
|
| Series | Advances in Neural Information Processing Systems |
| Event | 37th Conference on Neural Information Processing Systems (NeurIPS 2023) |
| Pages (from-to) | 61625-61644 |
| Number of pages | 20 |
| Publisher | Neural Information Processing Systems Foundation |
| Organisations |
|
| Abstract |
We consider the adversarial linear contextual bandit setting, which allows for the loss functions associated with each of K arms to changeover time without restriction. Assuming the d-dimensional contexts are drawn from a fixed known distribution, the worst-case expected regret over the course of T rounds is known to scale as Õ(√KdT). Under the additional assumption that the density of the contexts is log-concave, we obtain a second-order bound of order \tildeO(K√dVT) in terms of the cumulative second moment of the learner's losses VT, and a closely related first-order bound of order Õ(K√dL∗T) in terms of the cumulative loss of the best policy L∗T. Since VT or L∗T may be significantly smaller than T, these improve over the worst-case regret whenever the environment is relatively benign. Our results are obtained using a truncated version of the continuous exponential weights algorithm over the probability simplex, which we analyse by exploiting a novel connection to the linear bandit setting without contexts.
|
| Document type | Conference contribution |
| Language | English |
| Published at | https://papers.nips.cc/paper_files/paper/2023/hash/c2201e444d2b22a10ca50116a522b9a9-Abstract-Conference.html |
| Other links | https://doi.org/10.52202/075280 |
| Downloads |
NeurIPS-2023-first-and-second-order-bounds-for-adversarial-linear-contextual-bandits-Paper-Conference
(Accepted author manuscript)
|
| Permalink to this page | |
