AI Safety Issues can be categorized into five main classes including "Safe Exploration", "Scalable oversight", Avoiding “reward hacking” and “wire heading”, "Avoiding negative side effects" and "Robustness to distributional shift" (Amodei et al. (2016). Concrete Problems in AI Safety).
Can we categorize safety assurance of Evolutionary Optimisation Algorithms (e.g. Genetic Algorithm) as Safe Exploration?
Is there any approach for quantitative safety evaluation of such algorithms?
Thanks for your time and consideration.
I'd say these are classes to categorize the safety issues, not algorithms or families of algorithms.
For safety assurance, you'd need to address all issues relevant to your application.
Although safe exploration might be most important for Evolutionary Optimisation Algorithms, you'd still want to also avoid negative side effects.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence