Convergence and Error Bounds for Universal Prediction of Nonbinary Sequences.pdf

下载文档

0
0
约2.93万字
约 12页
2017-04-09 发布于江苏
举报
版权申诉
保障服务

Convergence and Error Bounds for Universal Prediction of Nonbinary Sequences.pdf

1、本文档共12页，可阅读全部内容。
2、有哪些信誉好的足球投注网站（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

Convergence and Error Bounds for Universal Prediction of Nonbinary Sequences

a r X i v : c s / 0 1 0 6 0 3 6 v 1 [ c s .L G ] 1 5 J u n 2 0 0 1 Technical Report IDSIA-07-01, 26. February 2001 Convergence and Error Bounds for Universal Prediction of Nonbinary Sequences Marcus Hutter IDSIA, Galleria 2, CH-6928 Manno-Lugano, Switzerland marcus@idsia.ch 1 http://www.idsia.ch/～marcus Keywords Bayesian sequence prediction; Solomonoff induction; Kolmogorov complexity; learn- ing; universal probability; finite non-binary alphabet; convergence; error bounds; games of chance; partial and delayed prediction; classification. Abstract Solomonoff’s uncomputable universal prediction scheme ξ allows to predict the next symbol xk of a sequence x1...xk?1 for any Turing computable, but otherwise un- known, probabilistic environment μ. This scheme will be generalized to arbitrary environmental classes, which, among others, allows the construction of computable universal prediction schemes ξ. Convergence of ξ to μ in a conditional mean squared sense and with μ probability 1 is proven. It is shown that the average number of prediction errors made by the universal ξ scheme rapidly converges to those made by the best possible informed μ scheme. The schemes, theorems and proofs are given for general finite alphabet, which results in additional complications as compared to the binary case. Several extensions of the presented theory and results are outlined. They include general loss functions and bounds, games of chance, infinite alphabet, partial and delayed prediction, classification, and more active systems. 1This work was supported by SNF grant 2000-61847.00 to Ju?rgen Schmidhuber. 1 Introduction The Bayesian framework is ideally suited for studying induction problems. The probability of observing xk at time k, given past observations x1...xk?1, can be computed with Bayes’ rule if the generating probability distribution μ, from which sequences x1x2x3... are drawn, is known. The problem, however, is that in many cases one does not even have a reasonable es