中文

用于桥接随机与对抗在线凸优化的乐观在线镜像下降

机器学习 2024-03-19 v3 机器学习

摘要

Sachs等人[2022]提出了随机扩展对抗(SEA)模型,作为随机与对抗在线凸优化之间的插值。在平滑条件下,他们证明了乐观跟随正则化领导者(FTRL)的期望遗憾依赖于凸函数的累积随机方差σ1:T2\sigma_{1:T}^2和累积对抗变差Σ1:T2\Sigma_{1:T}^2。他们还针对强凸函数给出了基于最大随机方差σmax2\sigma_{\max}^2和最大对抗变差Σmax2\Sigma_{\max}^2的稍弱界。受他们工作启发,我们研究了SEA模型下乐观在线镜像下降(OMD)的理论保证。对于凸且平滑函数,我们得到了相同的O(σ1:T2+Σ1:T2)\mathcal{O}(\sqrt{\sigma_{1:T}^2}+\sqrt{\Sigma_{1:T}^2})遗憾界,且无需单个函数的凸性要求。对于强凸且平滑函数,我们建立了O((σmax2+Σmax2)log(σ1:T2+Σ1:T2))\mathcal{O}((\sigma_{\max}^2 + \Sigma_{\max}^2) \log (\sigma_{1:T}^2+\Sigma_{1:T}^2))界,优于其O((σmax2+Σmax2)logT)\mathcal{O}((\sigma_{\max}^2 + \Sigma_{\max}^2) \log T)结果。对于指数凹且平滑函数,我们实现了新的O(dlog(σ1:T2+Σ1:T2))\mathcal{O}(d\log(\sigma_{1:T}^2+\Sigma_{1:T}^2))界。得益于OMD框架,我们拓展工作以研究动态遗憾最小化及在线函数非平滑的情形。我们建立了SEA模型下凸且平滑函数的首个动态遗憾保证,在非平稳场景中优于静态遗憾界。此外,为处理SEA模型中的非平滑凸函数,我们提出了基于带隐式更新的乐观OMD的新算法,可在无平滑条件下获得静态遗憾与动态遗憾保证。

关键词

引用

@article{arxiv.2302.04552,
  title  = {Optimistic Online Mirror Descent for Bridging Stochastic and Adversarial Online Convex Optimization},
  author = {Sijia Chen and Yu-Jie Zhang and Wei-Wei Tu and Peng Zhao and Lijun Zhang},
  journal= {arXiv preprint arXiv:2302.04552},
  year   = {2024}
}

备注

v3 substantially improves the presentation and has a few improvements, including the regret bound for strongly convex functions; v2 is an extended version that enriches the content with improved regret bounds for strongly convex functions, discussions on the optimism design for dynamic regret minimization, and extensions to non-smooth scenarios; v1 is the ICML 2023 conference version