CAMP: Counterexamples, Abstraction, MDPs, and Policy Refinement for Enhancing Safety, Stability, and Rewards in Reinforcement Learning
Year: 2025
Type: article
Abstract: Reinforcement learning (RL) has demonstrated exceptional performance across various real-world applications such as autonomous driving, robotic control, and finance. However, challenges surrounding safety and stability continue to limit its practical deployment. Specifically, effectively blocking transitions to failure states, maintaining consistent policy action selection, and optimizing rewards simultaneously remain key challenges in RL. This study addresses these issues by proposing a novel approach that combines counterexample-driven cluster refinement with a Markov Decision Process-Proximal Policy Optimization (MDP-PPO) hybrid policy. The proposed method enhances computational efficiency in complex state spaces through state abstraction and strengthens policy safety by employing the PRISM model checker to quantitatively analyze transitions to failure states. Furthermore, the approach is designed to optimize RL’s primary objective: reward maximization. Experimental evaluations in the CartPole, LunarLander, and Highway environments demonstrate significant improvements over the PPO-only baseline across all metrics. For example, in the LunarLander environment, the failure probability was reduced to 0.001352 (representing a reduction of over 90% relative to the baseline), while in the Highway environment, the standard deviation of rewards decreased by 85.27%, indicating a substantial gain in stability. In the CartPole environment, the average reward increased by 78.34%, reflecting the method’s ability to achieve superior performance in reward optimization. These quantitative results, combined with qualitative insights into enhanced policy safety and reliability, establish a novel framework for balancing safety, stability, and reward optimization in RL policy design. The integration of state abstraction and model checking plays a critical role in ensuring the reliability of policies, even in complex environments. Future work will focus on extending the approach to real-time applications and assessing its generalizability across diverse real-world scenarios.
Source: IEEE Access
Authors: Ryeonggu Kwon, Gihwon Kwon
Institution: Kyonggi University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W4409014295