Capturing Emergent Behavior In Multi-Response Systems Through Data Trend Mining
Conrad S. Tucker, Harrison Kim · 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference · 2010
This paper presents a novel approach to capture emerging systems behavior involving multiple performance criteria. Due to the interactions that exist among systems, engineers may be faced with a multi-objective design space that current single response data mining models do not capture. We aim to address this challenge by proposing a Multi-Response Trend Mining algorithm that simultaneously predicts multiple performance objectives by identifying the time series behavior of the individual systems. The proposed approach is a departure from conventional data mining approaches that are often limited to evaluating single response variables in a given static data set. The resulting system level predictions will serve as performance targets for next generation systems design eorts. The MultiResponse Trend Mining model can then be integrated with multi-objective engineering models during the systems design and analysis phase so that engineering design solutions better reect emerging system performance trends. A vehicle design data set from the UC Irvine Machine Learning Repository is used to validate the proposed methodology and highlight the need for multi-response predictive algorithms in systems design.