08:30 – 09:20 Introduction to Bayesian Optimization: Part I - Fundamentals 

09:20 – 10:10 Introduction to Bayesian Optimization: Part II - Hands-on Tutorial

10:10 – 10:40 Coffee Break

10:40 – 11:20 Bayesian Optimization for High-Dimensional, Adaptive, and Safe Controller Learning

11:20 – 12:00 Constrained Bayesian Optimization for Real-Time Control

12:00 – 13:30 Lunch Break

13:30 – 14:10 Multi-objective BO for Adaptive Embedded Predictive Control with Application to Personalized Plasma Medicine

14:10 – 14:50 Towards Self-Learning Industrial Process Plants: Leveraging Ideas from Bandit Problems and BO

14:50 – 15:20 Coffee Break

15:20 – 16:00 Bayesian Optimization Strategies for Real-World Manufacturing Efficiency

16:00 – 16:40 Bayesian Optimization and Autotuning in data-driven control: opportunities and challenges

16:40 – 17:00 Concluding Remark

08:30 -- 10:10  Introduction to Bayesian Optimization

Speaker: Yuning Jiang (EPFL) and Wenjie Xu (EPFL & Empa)

Abstract: Bayesian optimization is a derivative-free sample-efficient global optimization method based on surrogate modeling. In this part, we plan to introduce the fundamental concepts and popular algorithms for Bayesian optimization. Then, a hands-on tutorial will be given. The tentative software packages used in this tutorial are GPy, a Python-based Gaussian process package, and Ax, a Python-based BO platform. To fully benefit from this tutorial, the participants are encouraged to bring their laptops. We will first guide the audience to implement their own BO algorithm based on GPy, which includes selecting the kernel, building the GP model, implementing the acquisition function, and sampling the black-box functions. We will then use the implemented algorithm to solve a simple artificial toy example. We will then dive into more advanced usage of the Ax package by applying Ax to several demonstration problems. Tentative demos include tuning an inverted pendulum and a building thermal controller, both in simulation. The code base will also be made available to the workshop participants. 

10:10 – 10:40 Coffee Break

10:40 – 11:20 BO for High-dimensional, Adaptive, and Safe Controller Learning

Speaker: Sebastian Trimpe (RWTH Aachen)

Abstract: Optimizing controller parameters is ubiquitous in engineering, demanding precise tuning for performance and adaptability across diverse tasks.  Bayesian Optimization (BO) has emerged as a prime approach, offering a versatile framework for efficient and systematic parameter learning from few experimental trials.  In this talk, we present new BO algorithms tailored to confront challenges in learning control: (i) a fusion of BO and gradient methods designed for high-dimensional problems, (ii) event-triggered BO facilitating flexible adaptation to time-varying settings, and (iii) enhanced safe BO methods tackling critical gaps in safe learning.  In addition to introducing new algorithms and theory, the presentation will showcase the practical application of these innovations through experimental results on robots and cars. 

11:20 – 12:00 Constrained BO for Real-Time Control

Speaker: Dinesh Krishnamoorthy (TU Eindhoven) and Colin N. Jones (EPFL)

Abstract: This talk will explore the advanced use of Bayesian Optimization (BO) for real-time control, focusing on managing unknown constraints effectively. We will critique traditional constrained BO methods and introduce innovative variants, including strict safety-critical constraints and more flexible penalty-based approaches, alongside a novel method that integrates feedback control for constraint management without probabilistic models. Additionally, the session will cover a refined BO framework that combines grey and black box elements to incorporate prior knowledge into the optimization process, presenting an optimism-driven algorithm that achieves reliable convergence. This approach extends to solving distributed BO problems with guaranteed rates and introduces preferential measurements for capturing subjective preferences, demonstrated in a building's occupancy comfort optimization scenario.

14:10 – 14:50 Towards Self-Learning Industrial Process Plants: Leveraging Ideas from Bandit Problems and BO

Speaker: Mehmet Mercangöz (Imperial College London)

Abstract: The operation of industrial process plants is becoming increasingly complex, faced with growing demands for reduced environmental impact and enhanced profitability and flexibility. Concurrently, plant owners are grappling with a shortage of skilled engineers and operators. This trend is driving the search for more advanced automation systems, aiming to minimize reliance on human intervention during operations. A critical component in the hierarchy of industrial automation solutions is real-time optimization (RTO). RTO endeavours to identify economically optimal operating conditions, considering information such as product and raw material prices, and costs of auxiliary inputs such as electricity or fuels, while adhering to operational constraints. The proposed talk will introduce common RTO problem formulations and outline the prerequisites for endowing these algorithms with self-learning capabilities. Special emphasis will be placed on self-learning in scenarios where plant models are largely unknown, and where operation involves safety-critical constraints. We will present an RTO approach named ARTEO, inspired by the Upper Confidence Bound or UCB algorithm, which is designed to offer the discussed self-learning features. However, the talk will also address the drawbacks of ARTEO, specifically the scaling of computational demand and the challenges in satisfying constraints with high probability. To tackle the issue of scalability in computational demand, we suggest the use of parametric model ensembles. This approach estimates the uncertainty of model predictions by utilizing the statistics of predictions from the ensemble. For enhancing constraint satisfaction rates, we will explore two strategies: employing prior domain knowledge through the use of basis functions, and directly integrating known physical principles of the system (e.g., mass and energy balances) into the optimization problem using a bilevel programming approach. Throughout the talk, several examples will be provided to illustrate the strengths and limitations of these algorithms. Finally, we will conclude with a discussion on open research problems in this field and potential future developments.

14:50 – 15:20 Coffee break

15:20 – 16:00 Bayesian Optimization Strategies for Real-World Manufacturing Efficiency

Speaker: Alisa Rupenyan (Zurich Uni. for Applied Sciences) 

Abstract: Bayesian optimization (BO) has established itself as a suitable black-box optimizaiton for practical problems due to its versatility and data efficiency. It has proved efficient in both tuning controller parameters, and in serving as a control algorithm for optimization of manufacturing processes. This talk will delve into advanced BO methods designed for manufacturing challenges. Specifically, I will discuss batch-based BO, continuous (run-to-run) optimization, and risk-averse BO. These approaches tackle manufacturing challenges including data sparsity, variability and drift, and heteroscedasticity, while prioritizing safety and data efficiency. The algorithms are inspired by real-world applications including plasma spray coating, additive manufacturing, and precision motion stages. We'll cover their implementation, the challenges encountered, and showcase their efficiency through BO optimization case studies in these areas.

16:00 – 16:40 BO and Autotuning in Data-Driven Control: Opportunities and Challenges

Speaker: Valentina Breschi (TU Eindhoven)

Abstract: Data-driven control is gaining particular interest as an alternative to model-based design as it allows for mapping data onto a control law without undertaking a complete modeling step. However, all data-driven approaches depend upon several degrees of freedom, whose tuning requires significant human expertise and can severely impact the final performance of the controller. In this context, Bayesian Optimization can be the key to automatizing the choice of these crucial parameters, thanks to its capability of efficiently and effectively exploring the hyperparameter space without requiring designers to pre-set a search grid. Meanwhile, this approach is guided by some high-level performance requirements given by the designer, therefore allowing for the incorporation of expertise and priors into the tuning process. By introducing a unified framework for existing autotuning procedures in data-driven control, this talk spotlights how Bayesian Optimization, already widely adopted in model learning and successfully employed in tuning model-based control schemes, can be used for autotuning in data-driven control. At the same time, the talk discloses the limits and potential of this combination of techniques through the help of a set of numerical case studies.

16:40 – 17:00 Concluding Remark