Enhancing Regression-based Analytics for Addressing Applied Research Needs in Construction Engineering & Management (CEM)

回归分析的结果是简单的方程，足以代表现实世界系统中的建筑工程历史数据问题，以及与快速发展的计算（移动、社交、云）、传感器技术、建筑信息模型（BIM）基础参数化设计数据库和物联网相关的新兴“大数据”问题。然而，回归并没有能够跟上快速的技术进步和实际应用的需要。在现实世界中，问题可能是最令人难以置信的，并且数据通常包含噪声或缺失信息，而解决问题的方法期望计算简单，快速校准，直接解释推理逻辑，并且在新数据可用时易于保持最新状态。为了让用户接受并真正有效，CEM中基于数据的、分析驱动的决策支持系统的用户体验不能被认为是一个“黑盒”或需要大量的“试验和错误”。拟议的研究计划将加强基于线性回归的分析，以支持CEM的建模、预测和生产力和成本绩效的改进。在追求简单性的同时，研究将解决以下关键挑战：(1)如何提高基于线性回归分析的复杂性和智能性，以配合CEM中大多数复杂应用问题的“非线性”？(2)如何在保持模型精度的前提下，通过选择最主要的输入特征来简化高维回归方程？(3)如何通过解析表征模型预测误差来定义与点值预测相关的不确定性？最终目标是开发一个系统的、科学的框架，可以普遍应用于“突破和征服”现实世界的应用问题，从而能够为工程和管理专业人员提供及时、有效和定量的决策支持。所创造的新知识将大大丰富现有的澳电教育课程，包括研究生和本科生的定量教学方法。拟议的研究计划将培养高素质的人才，并提供改变游戏规则的解决方案，这将对行业产生重大影响。涉及数据驱动决策的其他相关领域也将从拟议的赠款中受益。

Regression analysis results in simple equations to sufficiently represent the real world systems in Construction Engineering historical data” problems as well as emerging “big data” problems in connection with rapid developments in computing (mobile, social, cloud), sensor technologies, parametric design databases underlying Building Information Models (BIM), and the Internet of Things. Yet, regression has not been able to catch up with rapid technology advances and practical application needs. In the real world, problems can be most mind-boggling, and the data often contain noises or missing information, while the problem-solving methods are expected to be computationally simple, fast to calibrate, straightforward to explain the reasoning logic, and easy to keep current as new data become available. In order to be acceptable and truly effective, user experiences of data-based, analytics-driven decision support systems in CEM must not be perceived as tapping a “black box” or requiring much “trial and error”.The proposed research program will enhance linear regression based analytics in support of modeling, prediction and improvement of productivity and cost performances in CEM. In parallel to the pursuit of simplicity, the research will address following crucial challenges: (1) how to enhance the sophistication and intelligence of linear-regression-based analytics so as to match up with the “non-linearity” native to most complicated application problems in CEM? (2) How to streamline high-dimensional regression equations by selecting the most dominant input features while retaining model accuracy? (3) How to define uncertainties associated with point-value predictions by analytically characterizing model prediction errors?The ultimate goal is to develop a systematic, scientific framework that can be generally applied to “break and conquer” real-world application problems, thus being capable to lend timely, effective, and quantitative decision support for engineering and management professionals in CEM. New knowledge to be created will substantially enrich existing CEM education curricula in regards to teaching quantitative methods on both graduate and undergraduate levels. The proposed research program will train highly qualified personnel along with delivering game-changing solutions that will make significant impact in the industry. Other related areas involving data-driven decision making will also benefit from the proposed grant.