A parameter-free hedging algorithm Export

@article{citeulike:4212873, abstract = {In this paper, we consider the decision-theoretic framework for online learning (DTOL) proposed by Freund and Schapire \cite{FS97}. Previous algorithms for learning in this framework have a tunable learning rate parameter. Tuning the learning rate requires prior knowledge about the sequence and severely limits the practicality of the algorithm. While much progress has been made in the past decade for adaptively tuning the learning rate, all of these methods still ultimately rely on some prior information. We propose a completely parameter-free algorithm for learning in this framework. We show theoretically that our algorithm has a regret bound similar to the best bounds achieved by previous algorithms with optimally-tuned learning rates. We also present a few experiments comparing the performance of the algorithm with that of other algorithms for various tunings.}, archivePrefix = {arXiv}, author = {Chaudhuri, Kamalika and Freund, Yoav and Hsu, Daniel}, citeulike-article-id = {4212873}, citeulike-linkout-0 = {http://arxiv.org/abs/0903.2851}, citeulike-linkout-1 = {http://arxiv.org/pdf/0903.2851}, day = {16}, eprint = {0903.2851}, keywords = {decision\_theory, nonparametric, online}, month = {Mar}, posted-at = {2009-03-24 06:24:49}, priority = {2}, title = {A parameter-free hedging algorithm}, url = {http://arxiv.org/abs/0903.2851}, year = {2009} }

TY - JOUR ID - citeulike:4212873 L3 - citeulike-article-id:4212873 N2 - In this paper, we consider the decision-theoretic framework for online learning (DTOL) proposed by Freund and Schapire \cite{FS97}. Previous algorithms for learning in this framework have a tunable learning rate parameter. Tuning the learning rate requires prior knowledge about the sequence and severely limits the practicality of the algorithm. While much progress has been made in the past decade for adaptively tuning the learning rate, all of these methods still ultimately rely on some prior information. We propose a completely parameter-free algorithm for learning in this framework. We show theoretically that our algorithm has a regret bound similar to the best bounds achieved by previous algorithms with optimally-tuned learning rates. We also present a few experiments comparing the performance of the algorithm with that of other algorithms for various tunings. TI - A parameter-free hedging algorithm KW - decision_theory KW - nonparametric KW - online AU - Chaudhuri, Kamalika AU - Freund, Yoav AU - Hsu, Daniel PY - 2009/Mar/16/ UR - http://arxiv.org/abs/0903.2851 ER -