Classic reinforcement learning (RL) is fundamentally online. The paradigm generally involves an agent interacting with its environment to gain experience; this experience is then later incorporated into the agent to improve future interactions.
Problem: Sometimes it is expensive or dangerous to learn in an online manner. It’s no good to have your driverless car speed off a cliff because it hasn’t figured out that cliffs are bad.
Solution: As in other areas of machine learning, it’d be ideal to throw a ton of data about past interactions in an environment at the RL algorithms and let them learn without (or with minimal) direct interaction with their environment. This data could ideally be simulated or gathered in some way that doesn’t require you to drive your autonomous car off the cliff.
The challenge, however, is reconciling this “stored data” vision (a.k.a. offline RL) with the online nature of many classic RL algorithms. The payoff to tackling this challenge will be to usher in a new class of “decision making engines” (in the same way that scaling up supervised learning ushered in a new class of “pattern recognition engines”).