Agent-based ecosystem modeling with deep reinforcement learning

Journal article, 2026

Ecosystem models can support understanding, monitoring, and management of ecological dynamics across space and time. Yet many existing animal-behavior simulations depend on manually specified rules, which limits scalability, transferability, and realism. We present a flexible agent-based modeling framework that leverages deep reinforcement learning to generate adaptive animal behavior without hand-coded decision rules. As a case study, we construct a model of an Alpine ecosystem comprising wolves, chamois, and vegetation, and evaluate it using Pattern-Oriented Modeling. The resulting simulations reproduce key ecological patterns, including long-term coexistence across multiple landscapes, predator-prey dynamics, and behavior qualitatively consistent with that of the modeled species. We further show how the model can be used to explore ecosystem resilience under scenarios of habitat degradation, game hunting, and heat stress. Finally, we compare our machine-generated model to a rule-based, hand-crafted model and observe that it outperforms the latter. While ecosystem modeling with deep reinforcement learning remains nascent and experimental, our approach provides a scalable step toward more flexible computational ecosystem models for exploring ecosystem responses to disturbance.