Cautious Bayesian Optimization: A Line Tracker Case Study.

Area of Science:

• Robotics and Control Systems
• Machine Learning and Artificial Intelligence
• Optimization Theory

Background:

• Experimental optimization often faces safety limitations.
• Traditional methods struggle with uncertainty and real-world constraints.
• Integrating safety into optimization is crucial for reliable system development.

Purpose of the Study:

• To present a novel procedure for experimental optimization under safety constraints.
• To enable safe fine-tuning of performance objectives despite experiment-model mismatch.
• To develop a robust optimization framework for complex systems.

Main Methods:

• Modeled performance and constraint functions using Gaussian processes.
• Incorporated transfer learning for prior mean modeling.
• Utilized a semi-parametric Kernel and chance-constrained acquisition function optimization.
• Developed a constraint-aware Bayesian Optimization (CABO) approach.

Main Results:

• Demonstrated safe experimental optimization through a case study.
• Successfully applied the methodology to a line-follower robot in CoppeliaSim.
• Validated the effectiveness of Gaussian process modeling for constraints.
• Showcased the ability to handle experiment-model mismatch safely.

Conclusions:

• Constraint-aware Bayesian Optimization provides a safe and effective method for experimental tuning.
• The proposed approach enhances reliability in optimizing systems with safety requirements.
• Gaussian processes and transfer learning are valuable tools for constrained optimization.