Uniform Error Bounds for Gaussian Process Regression with Application to Safe Control
Armin Lederer · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2020
Motivating Example Nonparameteric regresssion offers great promises in robotic applicationsLearned policies are unsafe in real world applications [1] • Constrained environments to avoid damages of hardware• No human-robot interaction due to risk of injuries START GOAL Quantification of uncertainty in data-driven models essential for safety-critical applications ⇒ Robust control for rigorous safety certificates How can the learning error be bounded based on the model uncertainty?How are formal safety guarantees provided for policies based on uncertain models? Gaussian Process Regression• Bayesian nonparametric modeling as "distribution over functions"