Augmenting Decision-Making in Complex and Safety-Critical Domains

This project, led by Assistant Professor Mark Crowley, focuses on problems of prediction and control in the areas of forest fire management, medical imaging and autonomous driving. The work on forest fires includes two main approaches. The first uses deep neural networks to learn compact models of forest fire spread directly from data such as satellite images or computationally expensive, physics based simulations. Another, more holistic approach taken uses reinforcement learning and game theory to learn a policy for wildfire spread across a landscape based on local conditions, as if the wildfire were an agent making decisions about where to move next. This approach utilizes multi-modal satellite, weather and other data to build more robust and generalizable models for prediction and decision making. These simulations are currently being applied to forest fire management but could apply to flood management, disease modelling and urban sprawl as well. This project involves collaborations with researchers in applied fields such as sustainable forest management, ecology, automotive technology and medical imaging.