Connecting Conservation Research and Implementation: Building a Wildfire Assistant

Sean McGregor, Rachel Houtman, Ronald Metoyer, Thomas G Dietterich

Abstract: Effective management of shared public ecosystems like national forests requires balancing beneficial uses (e.g., biodiversity and timber production). For forest wildfire suppression decisions, the balancing process tends to favor myopic policies that increase the risk of catastrophic fires through time. This work introduces WildFireAssistant as a visual analytics system combining a 100 year forest simulator with black box optimization. WildFireAssistant incorporates long-term impacts into decisionmaking by facilitating forest managers in their exploration of policies for different stakeholder groups. Since a critical feature for WildFireAssistant is the ability to interactively change the reward function (e.g., to explore the differences between policies optimized for biodiversity and timber), developing methods that support fast simulation and optimization are required. To this end, we develop a Model-Free Monte Carlo surrogate that pre-computes state transitions for fast visualization of policy functions. We then demonstrate a black box optimization method that utilizes the surrogate to support interactive optimization of user-selected reward functions.