Safe Exploration in Reinforcement Learning: Theory and Applications in Robotics
user-5ebe3bbdd0b15254d6c50b2c(2019)
摘要
Reinforcement learning has seen significant advances over the last decade in simulated or controlled environments. These successes have lead to interest in deploying learning algorithms in the real world, where they face significant prior uncertainties. While these algorithms are often able to find high-performance control strategies eventually, they typically do not provide any safety guarantees during the learning process. As a consequence, they cannot be deployed in safety-critical systems without posing a significant safety risk to both the learning system and its environment.In this dissertation, we introduce safe exploration algorithms that provide rigorous safety guarantees during the learning process. In particular, our algorithms explicitly model uncertainty about their environment in order to make safe decisions. These kind of algorithms are conservative in the beginning, when uncertainties are large, but become more confident over time as they acquire more data and learn about their environment. Importantly, they remain safe at all times during the learning process.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络