The nonparametric bandit approach to machine learning

This study is intended to survey results from a different path, best known amongst statisticians as bandit methods. These techniques also have an extensive history, and in recent years, have yielded new understanding. They are applicable to the same classes of problems as neural networks and genetic algorithms. The goal of this paper is to summarize some of the more recent developments with respect to nonparametric bandit theory. In particular, for many problem classes, upper bounds for convergence rates are known, regardless of learning method, and can be achieved by bandit techniques. Moreover, as documented here, there is a comprehensive theory applicable to the case in which noises are dependent. This is important to applications in networking and queues. The theory and case studies are cited in this work should be adequate to introduce the reader to principal developments, and serve as foundation for judgement on the relative merits of this machine-learning approach