Value-focused Robust Optimization for Disaster Response

The focus of the proposed project concerns logistics for responding to a disaster such as a hurricane, building on the results of a prior project by the PIs. Predominately, the paradigm in developing a decision-support system (DSS), including one for disaster response, is to first formulate a set of scenarios using available data or forecasts. Sometimes these are simply a handful of plausible scenarios, and other times the scenarios are derived from a more formal probabilistic model, which is fit to available data. Subsequently, these scenarios are fed into an optimization model, or a simulation-optimization model, which captures the “physics” of the underlying systems, which in our setting involve the transportation system and the liquid-fuel supply chain. The optimization model provides recommendations regarding a proactive response plan, e.g., how to deploy emergency responders, how to dispatch a fleet of diesel-fuel tank trucks, and how to advise populations evacuating. The proposed project will deviate from this main-stream paradigm by integrating the fitting of the probabilistic forecasting model and the DSS’s optimization model. The key driver behind the need for this integration is that in the approach sketched above, we first minimize the errors in our forecasting model. However, this process completely ignores which errors in the forecasting models make a difference in the optimization model. For example, a large error in the forecast for diesel fuel demand in a particular county is arguably unimportant if it does not have a significant impact on the logistics decisions we make, or more precisely a significant impact on the value of key performance metrics associated with these logistics decisions. Put simply, our goal is to focus on differences that make a difference, i.e., to focus on how the statistical errors in our forecasting models impact the quality of the decisions we recommend.