Validation of Dynamic Bayesian Optimization for a Non-Stationary Human-in-the-Loop Optimization Problem

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Human-in-the-Loop Optimization (HILO) is effective for assistive and augmentative tasks.
• Conventional Bayesian Optimization (BO) struggles with non-stationary human-robot systems, limiting applications in training and rehabilitation.

Purpose of the Study:

• To implement HILO using Dynamic Bayesian Optimization (DBO) for optimizing control parameters in robot-assisted training.
• To compare DBO against BO in a non-stationary environment simulating robot-assisted rehabilitation.

Main Methods:

• HILO was implemented using DBO and BO to determine optimal hip exoskeleton torques for maximizing hip extension angle during walking.
• Sixteen participants walked on a treadmill with gradually increasing speed to induce non-stationarity.
• The performance of DBO and BO was evaluated based on parameter convergence and accuracy.

Main Results:

• DBO demonstrated superior performance compared to BO, especially in later training stages.
• DBO led to a lower cost function and more accurate final torque values.
• DBO effectively modeled the history-dependent input-output relationship, outperforming BO's handling of process noise.

Conclusions:

• DBO is a more effective HILO strategy than BO for non-stationary human-robot systems.
• DBO's ability to model system dynamics improves performance in robot-assisted training and rehabilitation.
• This study highlights DBO's potential for adaptive and personalized robotic assistance.