TutORial: Coalescing Data and Decision Sciences for Analytics

By Suvrajeet Sen, Yunxiao Deng, and Junyi Liu.

The dream of analytics is to work from common data sources, so that all of its facets (descriptive, predictive, and prescriptive) are supported via a coherent data-driven vision. This vision of analytics is what we refer to as “Integrative Analytics”. In this tutorial we will cover a variety of OR/MS applications that require specific statistical learning models to be integrated with optimization models. For instance, certain cross-sectional data describing dependence among random variables may lead to regression models with multivariate error terms to be integrated with Stochastic Programming (SP) models. Others may require time series models to be integrated with Stochastic Model Predictive Control (S-MPC). Still other examples lead to particle filtering models providing data for network routing. In essence this tutorial will use these illustrations to motivate a new class of models, which we refer to as Learning Enabled Optimization (LEO) models. As suggested in the title of this tutorial, the applications are derived from integrative analytics. In addition to presenting these examples, the tutorial will cover fundamental concepts for modeling, statistically approximate solution concepts, sampling-based algorithms, and finally, model assessment and selection in the context of LEO models. Given the novelty of this paradigm, we will also outline how instructors may use the material for a graduate course on integrative analytics.