From Decision Support to Decision Services: An Expanded Role for AI in the Weather Enterprise

Artificial intelligence has long been an important tool for creating weather-influenced decision support systems, that is, algorithms, software and displays that provide information suitable to assist human forecasters and decision makers. For instance, AI effectively combines information from disparate sources, improves weather predictions, and translates weather conditions to human-relevant impacts. But in recent years, AI is increasingly proving able to go well beyond providing foundational data for decision-making: recent advances in applications as diverse as strategy games, internet advertising and self-driving cars show that AI can learn to quickly make optimal decisions in complex, dynamic environments. Since weather influences a huge number of decisions every day around the world – from taking an umbrella when leaving home given a chance of rain to packing up and evacuating when threatened by a hurricane – it is natural to ask: can the role of AI in the weather enterprise extend from decision support to providing (explainable) decision-making services across this spectrum?

This presentation describes a framework for weather-influenced decision optimization that suggests an affirmative answer. Key ingredients include (1) a calibrated ensemble of equally-likely forecasts that span the range of possible outcomes, called “prototypes,” which are derived via machine learning and statistical techniques from a diverse set of operational numerical weather prediction (NWP) models; (2) physical modeling or supervised learning to infer weather impacts from these multi-variate weather scenarios; (3) expert heuristic or machine-learned models for mapping actions to expected economic outcomes for a weather impact scenario; and (4) the use of cost/loss, n-armed bandit or reinforcement learning to compare alternative actions or policies and select the optimal one(s). This framework suggests that increasing the realized value of weather information to society may require improved methods for creating alternative realistic, multi-variate, spatiotemporal trajectories that span the possible weather outcomes; engagement with domain experts to understand and model how actions influence economic outcomes under disparate weather conditions for diverse personas and industries; and developing effective ways to communicate results and incorporate feedback to improve the decision system. Example applications of weather-influenced decision optimization using this framework will be described for several real-world use cases.