Search for: All records

Machine learning provides valuable information for data-driven decision-making. However, real-world problems commonly include uncertainties and the features needed to generate the prediction outputs are random variables. Even the most reliable machine learning models may not be helpful for decision-makers when the decisions must be taken before the values of features used in machine learning models are realized. To support decision-making under uncertainty, we propose a scenario generation procedure for stochastic programs that incorporates the uncertainties in both prediction features and the machine learning model prediction error. A statistical test is implemented to assess the reliability of the scenario sets by comparison with corresponding historical observations. We test the whole procedure in a case study for crop yield in Midwest. 

The mass transportation distance rank histogram (MTDRh) was developed to assess the reliability of any given scenario generation process for a two-stage, risk-neutral stochastic program. Reliability is defined loosely as goodness of fit between the generated scenario sets and corresponding observed values over a collection of historical instances. This graphical tool can diagnose over- or under-dispersion and/or bias in the scenario sets and support hypothesis testing of scenario reliability. If the risk-averse objective is instead to minimize CVaR of cost, the only important, or effective, scenarios are those that produce cost in the upper tail of the distribution at the optimal solution. We describe a procedure to adapt the MTDRh for use in assessing the reliability of scenarios relative to the upper tail of the cost distribution. This adaptation relies on a conditional probability distribution derived in the context of assessing the effectiveness of scenarios. For a risk-averse newsvendor formulation, we conduct simulation studies to systematically explore the ability of the CVaR-adapted MTDRh to diagnose different ways that scenario sets may fail to capture the upper tail of the cost distribution near optimality. We conjecture that, as with the MTDRh and its predecessor minimum spanning tree rank histogram, the nature of the mismatch between scenarios and observations can be observed according to the non-flat shape of the rank histogram. On the other hand, scenario generation methods can be calibrated according to uniform distribution goodness of fit to the distribution of ranks. 

Sustainable provision of food, energy and clean water requires understanding of the interdependencies among systems as well as the motivations and incentives of farmers and rural policy makers. Agriculture lies at the heart of interactions among food, energy and water systems. It is an increasingly energy intensive enterprise, but is also a growing source of energy. Agriculture places large demands on water supplies while poor practices can degrade water quality. Each of these interactions creates opportunities for modeling driven by sensor-based and qualitative data collection to improve the effectiveness of system operation and control in the short term as well as investments and planning for the long term. The large volume and complexity of the data collected creates challenges for decision support and stakeholder communication. The DataFEWSion National Research Traineeship program aims to build a community of researchers that explores, develops and implements effective data-driven decision-making to efficiently produce food, transform primary energy sources into energy carriers, and enhance water quality. The initial cohort includes PhD students in agricultural and biosystems, chemical, and industrial engineering as well as statistics and crop production and physiology. The project aims to prepare trainees for multiple career paths such as research scientist, bioeconomy entrepreneur, agribusiness leader, policy maker, agriculture analytics specialist, and professor. The traineeship has four key components. First, trainees will complete a new graduate certificate to build competencies in fundamental understanding of interactions among food production, water quality and bioenergy; data acquisition, visualization, and analytics; complex systems modeling for decision support; and the economics, policy and sociology of the FEW nexus. Second, they will conduct interdisciplinary research on (a) technologies and practices to increase agriculture’s contributions to energy supply while reducing its negative impacts on water quality and human health; (b) data science to increase crop productivity within the constraints of sustainable intensification; or (c) decision sciences to manage tradeoffs and promote best practices among diverse stakeholders. Third, they will participate in a new graduate learning community to consist of a two-year series of workshops that focus in alternate years on the context of the Midwest agricultural FEW nexus and professional development; and fourth, they will have small-group experiences to promote collaboration and peer review. Each trainee will create and curate a portfolio that combines artifacts from coursework and research with reflections on the broader impacts of their work. Trainee recruitment emphasizes women and underrepresented groups.