Using Dynamic Bayesian Optimization to Induce Desired Effects in the Presence of Motor Learning: a Simulation Study

Area of Science:

• Human-Computer Interaction
• Control Systems Engineering
• Computational Neuroscience

Background:

• Human-in-the-loop (HIL) optimization tunes parameters for human-interacting devices, but standard Bayesian Optimization (BO) assumes static user responses.
• Dynamic Bayesian Optimization (DBO) incorporates time into BO's kernel function to model changing user responses.

Purpose of the Study:

• To investigate if Dynamic Bayesian Optimization (DBO) is superior to standard Bayesian Optimization (BO) for HIL optimization when human responses exhibit motor learning.
• To evaluate DBO's effectiveness against BO using simulations based on human motor learning models.

Main Methods:

• Simulated human-in-the-loop optimization scenarios using both standard Bayesian Optimization (BO) and Dynamic Bayesian Optimization (DBO).
• Employed state-space models of human motor learning to simulate participant responses, including adaptation and use-dependent learning.
• Statistically compared the convergence performance of BO and DBO under various simulated response dynamics.

Main Results:

• Dynamic Bayesian Optimization (DBO) was consistently as good as or better than standard Bayesian Optimization (BO).
• DBO demonstrated superior convergence to optimal inputs and outputs after a specific number of iterations.
• The advantage of DBO emerged earlier in simulations with more dynamic user responses.

Conclusions:

• Dynamic Bayesian Optimization (DBO) presents a more effective paradigm than standard Bayesian Optimization (BO) for HIL optimization tasks involving human motor learning.
• DBO's improved performance is significant when sufficient iterations allow differentiation between true learning and random variability.