Machine Learning in Engineering Design: A Constructive Induction Approach

这项跨学科的研究旨在确定最近发展的机器学习方法在获取和/或创建概念设计知识方面的可行性和实用性。具体地说，与仅使用用户预先给出的属性的传统机器学习方法不同，本研究调查、增强并适应于设计问题的新的建构性归纳方法，这些方法能够自动地创造新的面向问题的属性。研究了两种新的方法--数据驱动的建构性归纳(DCI)和假设驱动的建构性归纳(HCI)。本研究的第一部分对高层建筑风支撑概念设计决策规则学习问题的方法进行了研究、改进和实验验证。主要训练数据是由设计和优化计算机系统产生的最优(最小重量)设计的集合。在第二部分，将利用所获得的方法论经验来建议和实施对方法的改进，并将它们协同地整合到一个专门针对概念设计问题的单一系统中。这一经验也将被用来开发一种在设计知识获取中应用建设性归纳学习系统的方法。由此产生的系统和方法将在工业环境中进行测试。这项研究有可能通过为设计者开发一种新的强有力的工具来显著改进概念设计，以及通过在实际和重要的现实世界问题的背景下进行研究来为机器学习领域做出贡献。

This interdisciplinary research aims at determining the feasibility and practicality of recently developed machine learning methods to the acquisition and/or creation of conceptual design knowledge. Specifically, this research investigates, enhances and adapts to the design problems new methods of constructive induction that are capable of automatically inventing new problem-oriented attributes, in contrast to the conventional machine learning methods that use only the attributes given in advance by the user. Two novel approaches are investigated-DCI (data-driven constructive induction) and HCI (hypothesis-driven constructive induction). The first part of this research studies, adapts, and experimentally tests the methods on the problems of learning decision rules for conceptual design of wind bracing in tall buildings. The primary training data is a collection of optimal (minimum weight) designs produced by a design and optimization computer system. In the second part, the methodological experience gained will be used to suggest and implement improvement to the methods, and to synergistically integrate them into a single system, specifically oriented toward problems of conceptual design. This experience will be also utilized to develop a method for applying constructive induction-based learning systems in design knowledge acquisition. The resulting system and method will be tested in an industrial environment. The research has a potential to significantly improve conceptual design by developing a new powerful tool for the designers, as well as to make contributions to the field of machine learning by conducting the study in the context of practical and important real-world problems.