The RxFire Recommendation Engine
My regional project for the CAP fellowship focuses on prescribed burning within the Southeast US. Opportunities for conducting prescribed burns are decreasing due to changing climate conditions and fluid operational parameters. We have developed a prototype recommendation engine which a) uses a Bayesian mixed effects model to calibrate 3-day weather forecasts used by fire managers to assess weather conditions and provides uncertainty estimates, and then b) combines that information with utility functions (e.g. years since last burn) to provide ranked recommendations of parcels to burn.
We ran a case study for the Eglin Air Force Base in Florida. Results indicate agreement between the recommendation engine and the observed decision, with the largest divergences likely arising primarily from differences between the presumed versus true utility function.