Comparison of Reinforcement Learning Approaches for Automated Control Derivation in Design Space Exploration for Safety-Critical Automotive Applications

This paper explores reinforcement learning for automated control derivation within design space exploration with focus on a functional safety concept for safety-critical automotive applications. A multi-task reinforcement learning framework is proposed to handle optimal control for various system topologies, component dimensioning, failures and scenarios. The timing analysis reveals that increasing the number of design variants significantly reduces per-topology training time, demonstrating the scalability of the proposed multi-task reinforcement learning approach for exploring large design spaces. This enables the derivation of optimal control across the entire design space, including both normal and failure conditions, while accounting for non linear plant dynamics with non-ideal actuator dynamics. The proposed methodology reduces manual engineering effort, supports derivation of fault tolerant control and offers a practical path toward automation in large-scale design space explorations.