Multistrategy Constructive Induction: A Theory and Methodology for Task-Oriented Improvement of Knowledge Representation Spaces for Learning

本研究涉及机器学习的一个基本问题，即在给定的表示空间质量不高的情况下如何进行学习。大多数机器学习方法在提供训练样本的相同表示空间中搜索期望的知识(即，使用与训练数据中使用的属性或术语相同的属性或术语)。如果原始表示空间不足，即给定的属性或描述性术语与当前学习任务不够相关，则这些方法不能很好地执行。这个项目调查了几种在这种情况下学习的新想法和方法。特别是，我们正在研究一种建设性的归纳方法来解决这个问题，在这种方法中，学习系统执行双重交织的搜索-一个是为了改进知识表示，另一个是为了在该空间中的“最佳”假设。在多策略建构性归纳法中，策略组合被用来将初始表征空间转换为新的面向任务的空间。空间变换操作符被分为通过发明新属性来扩展空间的“构造器”和通过移除或抽象当前属性来减少空间的“析构器”，其遵循最初提供的、但通过二次学习过程逐渐改进的“元规则”。

This research concerns a fundamental problem of machine learning, which is how to learn when the originally given representation space is of low quality. Most machine learning methods search for desirable knowledge in the same representation space in which training examples are presented (that is, use the same attributes or terms as those employed in the training data). If the original representation space is inadequate, that is, the given attributes or descriptive terms are insufficiently relevant to the present learning task, such methods do not perform well. This project investigates several novel ideas and methods for learning in such situations. In particular, we are investigating a constructive induction approach to the problem, in which a learning system performs a double intertwined search - one for an improved knowledge representation and one for the "best" hypothesis in that space. In the multistrategy constructive induction methodology, a combination of strategies is employed for transforming the initial representation space into the new task-oriented space. the space transformation operators are divided into "constructors" that expand the space by inventing new attributes and "destructors" that reduce the space by removing or abstracting the current attributes, guided by "meta rules" that are initially provided, but incrementally improved by a secondary learning process.