AI Agent Discovers and Controls Self-Organizing Patterns in Complex Systems

Traditional methods for exploring complex systems, such as cellular automata, typically involve setting initial conditions and observing the full simulation without real-time intervention. A new closed-loop framework, called CARL, introduces an "artificial experimentalist" agent based on autotelic reinforcement learning. This agent autonomously samples diverse goals and learns to intervene in complex systems using minimal, localized perturbations. Instantiated on Lenia, a continuous cellular automaton known for its life-like patterns, CARL demonstrates three key capabilities. First, it efficiently discovers stable solitons across various Lenia rules, outperforming heuristic methods. Second, it learns to precisely steer the movement of existing solitons with few interventions, showcasing its ability to control emergent patterns. Finally, CARL allows human users to guide solitons through maze environments in real-time by translating high-level directional commands into low-level interventions. The agents' policies, trained across diverse conditions, generalize effectively to out-of-distribution scenarios, paving the way for AI experimentalists that can autonomously or collaboratively discover and control complex emergent phenomena.