ABMax framework tunes AI agents to replicate classic Lotka-Volterra cycles

Researchers Corinna Mandl, Siddharth Chaturvedi, and Marcel van Gerven tackle a core challenge in complex adaptive systems: preventing runaway behavior or population collapse in agent-based models. Their paper presents ABMax, a JAX-based framework that enables efficient batched simulation of predator-prey ecosystems on hardware accelerators (GPUs/TPUs). Sheep and wolves are modeled as active agents with local sensing, internal energy, and recurrent neural network-based controllers. The team optimizes environmental and demographic parameters using a feature-based loss that rewards sustained oscillations, phase lag, bounded populations, and long-term persistence—first for random controllers, then for evolved ones in more naturalistic settings.

This approach successfully nudges agent-based dynamics toward classical Lotka-Volterra cycles, demonstrating that carefully tuned local rules can yield predictable, stable global patterns. The use of JAX allows massive parallelization, making large-scale tuning feasible. For professionals working in multi-agent AI, robotics, or ecological modeling, ABMax offers a principled method to stabilize simulations and bridge the gap between emergent behavior and theoretical models. The paper (12 pages, 3 figures) is available on arXiv (2606.13639).