Study on Pattern Inference from Positive Data

本研究的目的是通过研究归纳推理作为示例学习的理论模型来考察机器学习的可实现性。一般来说，学习中使用的例子分为积极的例子和消极的例子。在语言(或语法)学习中，积极的例子对应于(语法上)正确的句子。从实验中获得的数据可以被认为是某一性质的正例，当它们与某一性质有关时。在本研究中，我们考虑了基于正例的归纳学习的理论局限性，从实际应用的角度研究了高效的学习算法。模式是由常量符号和变量组成的字符串。模式的语言是由获取的常量字符串的集合。用非空常量字符串替换模式中的变量。对于任何固定的k，已经证明了至多k个模式语言的并类可以从正数据中推断出来，我们应用模式语言的学习算法从氨基酸序列中发现一个基元。在字母索引的帮助下，算法成功地从正例中找到了可以被认为是模体的模式集。我们还研究了基本形式系统中语言接受者的加速问题，其中我们使用了快速的字符串模式匹配机。最后，我们提出了一种扩展多模式学习算法的可能方法。

The aim of this research is in investigating realizability of machine learning, by studying inductive inferece as a theoretical model of learning from examples. In general, examples using in learning are categorized in positive ones and negative ones. In language (or grammar) learning, positive examples are corresponding to (grammatically) correct sentences. Data obtained from experiments can be considered as positive examples of a certain property, when they are concernd with the property. In this research, we have considered theoretical limits of inductive learning based on positive examples and investigated efficient learning algorithms from the viewpoint of practical applications.A pattern is a string consisting of constant symbols and variables. The language of a pattern is the set of constant strings obtained by. substituting nonempty constant strings for variables in the pattern. For any fixed k, the class of unions of at most k pattern languages is already shown to be inferable from positive data.We apply a learning algorithm for pattern languages to discover a motif from amino-acid sequences. From only positive examples with the help of an alphabet indexing, the algorithm successfully finds sets of patterns, that can be considered as motifs.We have also studied speed-up of language acceptors for elementary formal systems, where we employ fast string pattern matching machines. Finally, we propose a possible approach to extending leaning algorithms for multiple patterns.