Combining off-Policy and on-Policy Reinforcement Learning for Dynamic Control of Nonlinear Systems

This paper introduces QARSA, a novel reinforcement learning algorithm that combines the strengths of off-policy and on-policy methods, specifically Q-learning and SARSA, for the dynamic control of nonlinear systems. Designed to leverage the sample efficiency of off-policy learning while preserving the stability and lower variance of on-policy approaches, QARSA aims to offer a balanced and robust learning framework. The algorithm is evaluated on the CartPole-v1 simulation environment using the OpenAI Gym framework, with performance compared against standalone Q-learning and SARSA implementations. The comparison is based on three critical metrics: average reward, stability, and sample efficiency. Experimental results demonstrate that QARSA outperforms both Q-learning and SARSA, achieving higher average rewards, stability, sample efficiency, and improved consistency in learned policies. These results demonstrate QARSA’s effectiveness in environments were maximizing long-term performance while maintaining learning stability is crucial. The study provides valuable insights for the design of hybrid reinforcement learning algorithms for continuous control tasks.