Reinforcement Learning with Guarantees that Hold for Ever

Abstract

Reinforcement learning is a successful explore-and-exploit approach, where a controller tries to learn how to navigate an unknown environment. The principle approach is for an intelligent agent to learn how to maximise expected rewards. But what happens if the objective refers to non-terminating systems? We can obviously not wait until an infinite amount of time has passed, assess the success, and update. But what can we do? This talk will tell.