Guided Bayesian Optimization: Data-Efficient Controller Tuning with Digital Twin

TL;DR

The paper tackles the challenge of data-efficiently tuning closed-loop controllers for uncertain or hard-to-model plants by framing the problem as a black-box optimization of a cost $J(\theta)$. It introduces Guided Bayesian Optimization, which combines Gaussian-process-based BO with an event-triggered digital twin to shift exploration from the real plant to a cheap DT when the surrogate uncertainty is high. The method is model-free, independent of plant/controller structure, and validated on two real hardware systems (a DC rotary motor and a linear servo motor), showing substantially fewer real experiments needed to reach near-optimal controller parameters compared with conventional BO. The work demonstrates practical gains in data efficiency and operational cost, with potential extensions to constrained, multi-objective, and online-DT-validating settings for broader industrial applications.

Abstract

This article presents the guided Bayesian optimization algorithm as an efficient data-driven method for iteratively tuning closed-loop controller parameters using an event-triggered digital twin of the system based on available closed-loop data. We define a controller tuning framework independent of the controller or the plant structure. Our proposed methodology is model-free, making it suitable for nonlinear and unmodelled plants with measurement noise. The objective function consists of performance metrics modeled by Gaussian processes. We utilize the available information in the closed-loop system to identify and progressively maintain a digital twin that guides the optimizer, improving the data efficiency of our method. Switching the digital twin on and off is triggered by data-driven criteria related to the digital twin's uncertainty estimations in the BO tuning framework. Effectively, it replaces much of the exploration of the real system with exploration performed on the digital twin. We analyze the properties of our method in simulation and demonstrate its performance on two real closed-loop systems with different plant and controller structures. The experimental results show that our method requires fewer experiments on the physical plant than Bayesian optimization to find the optimal controller parameters.

Figures (15)

• Figure 1: Efficient data-driven controller tuning scheme based on performance assessment
• Figure 2: Guided BO schematic representation. The digital twin of the closed-loop system is built using available data without additional operations on the real system.
• Figure 3: Details of the performance metrics used to define the overall cost function
• Figure 4: Results for different activation threshold $\eta_{1}$ of DT in guided BO method. Each data point averages over $100$ batches, each including $50$ experiments on the real plant. (top) The required number of experiments on the real system to outperform the nominal controller performance or to converge to the ground truth performance $J(\theta^{*})$. The shaded area indicates $\eta_{1}$ range where guided BO converges faster to the ground truth performance than BO. (bottom) The left axis is the average optimality ratio. The right axis shows the average number of DT activations in each batch.
• Figure 5: Minimum observed optimality ratio up to number of BO experiments on the real plant. The thick line is the average over $100$ batches, and the shaded area shows the $99\%$, $95\%$, $90\%$, and $68\%$ confidence intervals.