Using dynamic Bayesian optimization to induce desired effects in the presence of motor learning: a simulation study

We sought to establish whether dynamic Bayesian optimization (DBO) is a suitable algorithm for human-in-the-loop-optimization (HILO) of the control input of devices interacting with individuals whose output changes during optimization as resulting from motor learning. Simulations were conducted assuming either purely time-dependent participant responses, or assuming responses from state-space models of motor learning. DBO generally outperformed standard Bayesian optimization (BO) in convergence to optimal inputs and outputs after a certain number of iterations. DBO may improve the performance of HILO over BO when a sufficient number of iterations can be evaluated to accurately distinguish between unstructured variability and learning.

本研究旨在探明动态贝叶斯优化（dynamic Bayesian optimization, DBO）是否适用于人机交互设备控制输入的人机循环优化（human-in-the-loop-optimization, HILO）——该优化场景中，设备的交互对象的输出会因运动学习过程在优化期间发生变化。本研究开展了两类仿真实验：一类假设参与者的响应仅随时间变化，另一类则采用运动学习的状态空间模型生成参与者响应。在经过一定数量的迭代后，动态贝叶斯优化在收敛至最优输入与输出的表现上，整体优于标准贝叶斯优化（Bayesian optimization, BO）。当可通过足够次数的迭代精准区分无结构变异性与运动学习过程时，动态贝叶斯优化相较于标准贝叶斯优化，可提升人机循环优化的性能。