The recent developments in the research area focus on enhancing the safety, efficiency, and optimization of autonomous systems and control mechanisms, particularly through advanced Bayesian optimization techniques and safety verification methods. A significant trend is the application of Bayesian optimization (BO) in various domains, from autonomous vehicle control to building temperature management, aiming to automate and optimize parameter tuning processes. Innovations in BO methods, such as early stopping, co-learning, and generative multi-form optimization, are addressing the challenges of sample efficiency, safety constraints, and complex input spaces. Additionally, there is a strong emphasis on ensuring the safety of autonomous systems, with novel approaches to decompose and quantify safety requirements for perception systems and to train neural networks with provable safety guarantees using hybrid zonotope reachability analysis.

Noteworthy Papers

• Decomposition and Quantification of SOTIF Requirements for Perception Systems of Autonomous Vehicles: Introduces a risk decomposition methodology to derive quantitative safety requirements for autonomous vehicle perception systems, facilitating safety verification.
• Which price to pay? Auto-tuning building MPC controller for optimal economic cost: Proposes an efficient method for tuning building MPC controllers using constrained Bayesian optimization, significantly reducing electricity costs.
• Early Stopping Bayesian Optimization for Controller Tuning: Develops a BO method that stops episodes early if suboptimality is detected, reducing optimization time.
• Lipschitz Safe Bayesian Optimization for Automotive Control: Presents a BO algorithm that satisfies multiple safety constraints simultaneously, applied to tuning a self-driving car's trajectory-tracking controller.
• Provably-Safe Neural Network Training Using Hybrid Zonotope Reachability Analysis: Offers a fast method for training neural networks to avoid unsafe regions, using hybrid zonotopes and MILPs.
• Co-Learning Bayesian Optimization: Introduces a novel BO algorithm that improves surrogate accuracy with limited samples by exploiting model diversity and agreement.
• Generative Multi-Form Bayesian Optimization: Proposes a multi-form GMO approach for optimizing over complex structured input spaces, enhancing solution accuracy and convergence rate.
• Safety in safe Bayesian optimization and its ramifications for control: Identifies practical safety obstacles in SafeOpt-type algorithms and proposes LoSBO and LoS-GP-UCB as solutions.