Integrating Explanation-Based and Neural Approaches to Machine Learning

将基于解释的方法和神经方法集成到机器学习中

• 批准号: 9002413
• 负责人: Jude Shavlik
• 金额: $ 17.01万
• 依托单位: University of Wisconsin-Madison
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1990
• 资助国家: 美国
• 起止时间: 1990-09-01 至 1994-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9002413&HistoricalAwards=false
• 关键词: Integrating; Explanation; Based; Neural; Approaches

To be considered intelligent, machines must be capable of learning. Symbolic and neural approaches to machine learning both have been heavily investigated. However, there has been little research into the synergies achievable by combining these two learning paradigms. This work will develop a hybrid system that combines the symbolically-oriented explanation-based learning paradigm with the neural back-propagation algorithm. In this system, the initial neural network configuration is determined by the generalized explanation of the solution to a specific classification of planning task. This research addresses the problem of choosing a good initial neural network configuration and overcomes problems that arise when using imperfect theories to build explanations. Most real-world problems can never be formalized exactly. However, there is much to be gained by utilizing the capability to reason approximately correctly. Explanation-based learning provides a way to profitably use casual models of the world, while neural networks provide a way to refine roughly-correct concepts. This research on combining these two learning paradigms promises to broaden the applicability of machine learning techniques, producing learning algorithms that are not brittle and which can produce concepts whose accuracy improves through experience.

要被认为是智能的，机器必须具有学习能力。 机器学习的符号和神经方法都经过了深入研究。 然而，很少有人研究通过结合这两种学习范式可以实现的协同作用。 这项工作将开发一种混合系统，将面向符号的基于解释的学习范式与神经反向传播算法相结合。 在该系统中，初始神经网络配置是通过对规划任务的特定分类的解决方案的广义解释来确定的。 这项研究解决了选择良好的初始神经网络配置的问题，并克服了使用不完美的理论建立解释时出现的问题。 大多数现实世界的问题永远无法准确地形式化。 然而，利用正确近似推理的能力可以获得很多好处。 基于解释的学习提供了一种有利地使用世界随意模型的方法，而神经网络提供了一种改进大致正确的概念的方法。 这项结合这两种学习范式的研究有望扩大机器学习技术的适用性，产生不易损坏的学习算法，并且可以产生通过经验提高准确性的概念。