Abstract: Forecasting the behavior of complex ecological systems is not merely a hard problem, but often impossible. Moreover, the model that provides the greatest predictive accuracy is not necessarily the model that yields the best decision. How then can scientific research guide conservation decision-making in face of such a complex and uncertain world? We often imagine the role of science as reducing uncertainty about the future using models and data, while leaving decision making about actions to advocates and policymakers. I will argue that we must turn this view on its head. Research develops theory and models which show possibilities -- scenarios -- that could arise, seeking to broaden, not narrow, our understanding of uncertainty. I will then show how the science of decision theory can act as a guide to find robust strategies in the face of such unknowns. I will illustrate how emerging methods in deep reinforcement learning (RL), used to train artificial intelligence systems in everything from chess to chatbots, can be applied to the search for robust strategies. Further, I will highlight how these remarkable algorithms are in fact re-discovering and implementing principles of adaptive management known to practioners of ecosystem management for many decades, aside some new (and sometimes concerning) tricks. I will also outline theoretical basis as to why this approach to robust decision-making may be more practical than predict-and-perscribe science.
