Formalising performance guarantees in meta-reinforcement learning

Abstract

Reinforcement learning has had great empirical success in different domains, which has left theoretical foundations, such as performance guarantees, lagging behind. The usual asymptotic convergence to an optimal policy is not strong enough for applications in the real world. Meta learning algorithms aim to use experience from multiple tasks to increase performance on all tasks individually and decrease time taken to reach an acceptable policy. This paper proposes to study the provable properties of meta-reinforcement learning.