Improving Stability and Efficiency of Reinforcement Learning using Reward Functions based on Lyapunov Neural Networks

Reinforcement Learning (RL) has demonstrated remarkable success but faces challenges in stability and safety. This paper introduces DroQ_LyNN that integrates Lyapunov Neural Networks into the Doubly Efficient Reinforcement Learning (DroQ) framework. By leveraging Lyapunov stability theory, the proposed method ensures safe and stable control while maintaining the efficiency of DroQ. We design a data-driven Lyapunov neural networks to estimate safe regions (ROA) and embed these into the reward function, enabling stable policy learning without compromising exploration. Experiments on OpenAI safety benchmarks show DroQ_LyNN outperforms baselines like DroQ, TD3, and DDPGLag in learning efficiency and control stability. This work highlights the potential of combining Lyapunov stability with RL for robust, safe learning in complex environments.The project is available at: https://github.com/needtorch/security-restrictions-LyNN.