Efficient Design Space Exploration with Multi-Task Reinforcement Learning

Exploring the design space is a critical aspect of engineering and optimization, involving the search for the best configuration in complex systems with numerous options. In the system design process, it is essential to take into account a range of constraints related to architecture and component dimensioning, as well as requirements defined by standards or the current state-of-the-art. One of the main challenges in design space exploration is developing a control strategy tailored to each specific design, facilitating an objective comparison of different designs for closed-loop scenarios. Even though reinforcement learning offers promise as an automated solution for deriving control strategies, its trial-and-error methodology demands significant computational resources. To address this challenge, leveraging knowledge from similar design combinations, especially in larger design spaces, becomes beneficial. This study specifically targets the speed-up of automated derivation of control strategies within design space exploration using multi-task reinforcement learning. The work is applied to a safety-critical cross-domain motion system, comprising drive, brake, and steer systems. It further considers different driving scenarios and failure cases, enabling system performance assessment in normal and various failure modes within a limited time frame. With the proposed speed up of automated derivation of control strategies the overall effectiveness of design space exploration for multi-actuated and integrated system architectures is enhanced.