Ensuring the Quality of Optimization Solutions in Data Generated Optimization Models

Abstract

Mathematical optimization models can improve decision making in a wide variety of domains, both industrial and societal. However, creating an optimization model requires rare optimization expertise and a significant amount of time, thereby limiting the widespread use of this technology. One way to overcome this is to enable data driven generation of complete optimization models from historical data. A major challenge posed by such data-driven generation is that the uncertainties and inaccuracies resulting from the model generation must be explicitly accounted for to ensure the quality of the solutions ultimately produced by the generated optimization model. Moreover, the historical data used for such model generation must also contain decision related data, which has very different characteristics from the input typically used for machine learning models, further compounding the difficulty of accounting for the uncertainties. In this work, it is our goal to bring to light this important topic of end-to-end data driven optimization model generation and encourage additional research in this area. Our contributions here are therefore to formally define the problem and outline several promising approaches for addressing it. We also describe areas for future work, which gives some indication of the breadth and depth of topics to be addressed. We hope this will also motivate others to carry out research in the broader field of enabling widespread creation of optimization models by people who are not optimization experts.