מושגי ליבה
Using heuristics for uncertainty bounds in SafeOpt algorithms can lead to safety violations, highlighting the importance of implementing Real-β-SafeOpt with theoretically sound choices.
תקציר
This analysis delves into the practical safety implications of using heuristics in frequentist uncertainty bounds within SafeOpt algorithms. It discusses the issues with heuristics, demonstrates safety problems through experiments, and proposes Real-β-SafeOpt as a solution. The study also introduces Lipschitz-only Safe Bayesian Optimization (LoSBO) to address assumptions related to RKHS norms and safety guarantees.
-
Introduction
- Safety constraints in optimization tasks.
- Importance of theoretical safety guarantees.
-
Background
- Gaussian Process regression and RKHS.
- Frequentist uncertainty bounds for GP regression.
-
Problem Setting and Objectives
- Investigating practical safety issues in SafeOpt.
- Proposing Real-β-SafeOpt as a solution.
-
Real-β-SafeOpt
- Using modern uncertainty bounds numerically.
- Implementing Real-β-SafeOpt algorithm.
-
Lipschitz-only Safe Bayesian Optimization (LoSBO)
- Addressing assumptions on RKHS norms.
- Ensuring verifiable and reasonable safety guarantees.
סטטיסטיקה
"2727±3882 of these runs (all 10000 repetitions for all 100 functions) lead to a bound violation."
"2862 out of 10000 runs with safety violations."