Reinforcement learning to influence the emerging behaviors of complex systems

This thesis investigates the control of simulated agent-based models using reinforcement learning (RL). Self-organizing systems, where interacting individuals exhibit complex and emergent patterns, can undergo phase transitions based on parameter adjustments. This research proposes a method to control the global dynamics of these systems by introducing RL agents that influence group behavior. These agents will be trained to learn a policy that guides them towards achieving a desired stable state for the overall system.