Safety-Counter-Player: Utilizing potentially unsafe capabilities in safety-critical systems

In safety-critical systems, integrating machine learning components (MLCs) presents significant challenges in balancing safety with functional performance. Engineers strive to harness machine learning to enhance both system functionality and safety. However, they face obstacles in ensuring sound safety assurance for these machine learning components. This paper proposes a novel architecture that distinguishes between two roles: the safety-player, which is responsible for making critical safety interventions, and the counter-player, which focuses on optimizing functional performance. By permitting the safety-player to intervene only when absolutely necessary, the counter-player is allowed greater freedom in its operations. This separation not only improves performance but also maintains safety, fostering a more effective interaction between safety, comfort, and overall system utility.