Bayesian Optimization Finds Energy-Efficient RL Policies Faster

Industrial automation increasingly demands control strategies that not only perform well but also adhere to strict energy efficiency standards. In reinforcement learning (RL), this often means combining competing objectives into a single reward function, a process that typically relies on manual weight selection. This manual approach is time-consuming, prone to bias, and frequently misses optimal trade-off solutions. This research tackles the critical problem of automating this weight selection to efficiently identify the Pareto front of optimal trade-off policies. The authors reframe the weight selection as a multi-objective Bayesian optimization (MOBO) problem. Using a physical Quanser Aero 2 testbed for 1-DoF pitch control, the study compared the MOBO approach, specifically with the expected hypervolume improvement (qEHVI) acquisition function, against a standard uniform grid search. Results consistently showed that MOBO achieved superior hypervolume and maximum spread, successfully identifying diverse, high-quality trade-off policies with a significantly reduced evaluation budget. This makes energy-aware control in complex mechatronic systems much more efficient.