Reinforcement Learning for Structural Control


This study focuses on improving structural control through reinforcement learning. For the purposes of this study, structural control involves controlling the shape of an active tensegrity structure. Although the learning methodology employs case-based reasoning, which is often classified as supervised learning, it has evolved into reinforcement learning, since it learns from errors. Simple retrieval and adaptation functions are proposed. The retrieval function compares the response of the structure subjected to the current loading event and the attributes of cases. When the response of the structure and the case attributes are similar, this case is retrieved and adapted to the current control task. The adaptation function takes into account the control quality that has been achieved by the retrieved command in order to improve subsequent commands. The algorithm provides two types of learning: reduction of control command computation time and increase of control command quality over retrieved cases. Results from experimental testing on a full-scale active tensegrity structure are presented to validate performance.