Abstract
Deep Reinforcement Learning (RL) applied to data-driven control exhibits promising outcomes for control tasks with continuous action spaces. A notable drawback of Deep RL lies in its computational demands, which pose a concern when considering deploying Deep RL applications on resource-constrained unmanned platforms. Research in mixed numerical precision methods is actively contributing to enhancing the computational efficiency of Deep Learning approaches. While mixed-precision strategies are established for supervised learning tasks, exploration of this realm is limited for Deep RL. Our objective is to bridge this gap in research by enhancing the computational efficiency of the Deep Deterministic Policy Gradient (DDPG) algorithm through the utilization of mixed-precision and loss scaling. We will present numerical cases to quantify the performance and computational advancements of DDPG agents trained with mixed precision in the context of control for a complex dynamic system model.