A Constraint-based Safety Model for Reinforcement Learning

Safe reinforcement learning is an emerging research area that focuses on developing algorithms and techniques to train reinforcement learning agents to act safely in real-world environments. While reinforcement learning has achieved remarkable success in various applications, using these agents in safety-critical systems such as self-driving cars, medical devices, and robotics poses significant challenges. The primary concern is that these agents may exhibit unsafe or unpredictable behavior, which can lead to severe consequences. Therefore, the goal of safe reinforcement learning is to develop methods that ensure the agent's behavior is safe and reliable in the face of uncertainty and unexpected conditions. This paper introduce CARL (Constraint Acquisition Reinforcement Learning) that is a framework for automatic and integrated constraint identification in reinforcement learning problems. CARL automatically identifying constraints from the agent's experiences and using them to guide the learning process towards safe and effective policies. Experimental results have shown that CARL can effectively learn policies that satisfy safety constraints in complex environments and can outperform traditional reinforcement learning algorithms that do not consider constraints.The CARL algorithm rapidly achieves maximum rewards with significantly fewer steps. However, it's observed that standard reinforcement learning algorithms, given more steps and time, can achieve higher maximum rewards.