Reinforcement Learning Reveals Spatial Strategies for Slowing Biological Invasions

Abstract

Managing biological invasions requires deciding where, when, and how to act while populations continue to grow and spread. Invasive species threaten ecosystems and livelihoods, but limited management resources mean that managers cannot treat every affected area. This study asks how control efforts can be directed across a landscape to slow an invasion while keeping costs low. We develop a reinforcement-learning framework that can use broader spatial context to control biological invasions, using the spongy moth as a case study. The stochastic simulator captures host-dependent growth, short- and long-distance dispersal, density-dependent mating, and the effects of multiple treatments. The reinforcement-learning strategies are compared with benchmarks that are based on local treatment rules. We find that reinforcement learning achieves the lowest mean cost, with its advantage widening as management complexity increases, such as with simultaneous advancing-front control, nascent-colony management, and established-colony containment on heterogeneous-host landscapes. The reinforcement-learning strategies exhibit highly context-dependent treatment behavior, adjusting treatment choices as invasion dynamics evolve and coordinating interventions across long distances, all of which is learned from scratch.