This thesis investigates optimization and reinforcement learning when useful signals, particularly rewards, are rare. It develops methods based on interacting particle systems and large deviation theory to improve exploration, prevent excessive diversification, and construct more reliable estimators using prior information.