Safe and Efficient Model-free Adaptive Control via Bayesian Optimization

Abstract

Adaptive control approaches yield highperformance controllers when a precise system model or suitable parametrizations of the controller are available. Existing data-driven approaches for adaptive control mostly augment standard model-based methods with additional information about uncertainties in the dynamics or about disturbances. In this work, we propose a purely data-driven, model-free approach for adaptive control. Tuning low-level controllers based solely on system data raises concerns on the underlying algorithm safety and computational performance. Thus, our approach builds on GOOSE, an algorithm for safe and sample-efficient Bayesian optimization. We introduce several computational and algorithmic modifications in GOOSE that enable its practical use on a rotational motion system. We numerically demonstrate for several types of disturbances that our approach is sample efficient, outperforms constrained Bayesian optimization in terms of safety, and achieves the performance optima computed by grid evaluation. We further demonstrate the proposed adaptive control approach experimentally on a rotational motion system.

Design and stability analysis of semi-implicit cascaded proportional-derivative controller for underactuated cart-pole inverted pendulum system

“…The considerations behind are that testing real systems with un-verified sets of parameters is unsafe, which may damage the mechanics and bring danger to human operators. Besides, fewer times of runs means less wear and tear of the system [15]. Therefore, BO is suitable to be integrated into the design procedures of our proposed algorithms.…”

Machine Learning for automatic pointing alignment and spatial beam filtering

“…We implemented a data-driven optimization in order to robustly find the optimal operation point within a few minutes. This machine-learning method, using Bayesian optimization [2] in a Python framework [3], was characterized on a laboratory setup utilizing a heliumneon laser, and a capillary spatial filter, as shown in Fig 1 . In this setup, the beam pointing direction and angle to the spatial filter are controlled by utilizing two motorized mirror mounts equipped with stepper motors and encoders. The beam position and the angle to the spatial filter are evaluated by running a Gaussian fit of the beam profile images provided by Cameras 1 and 2.…”

On Controller Tuning with Time-Varying Bayesian Optimization

“…The time-varying UCB variant increases the UCB exploration parameter over time, which leads to undesirable high exploration in controller tuning problems and is difficult to tune in practice. The work of Bogunovic et al [9] has been adapted to different applications in low-dimensional settings with up to four parameters, such as controller learning [17], safe adaptive control [18], and online hyperparameter optimization [19]. In contrast to the algorithms based on Bogunovic et al [9], we use the standard UCB algorithm for TVBO without increasing exploration over time.…”

“…However, controller tuning with changing dynamics is significantly less explored in literature. König et al [18] considers an adaptive control problem for safe model-free adaptive control where they adopt the setting of Bogunovic et al [9]. Shape Constraints in BO: The main benefit of incorporating prior knowledge in BO is an increase in sample efficiency as it reduces the hypothesis space of the objective function.…”