Reinforcement Learning

Chapter 10. Operational Reinforcement Learning

The closer you get to deployment, the closer you are to the edge. They call it the cutting edge for good reason. It’s hard enough getting your reinforcement learning (RL) project to this point, but production implementations and operational deployment add a whole new set of challenges.

To the best of my knowledge, this is the first time a book has attempted to collate operational RL knowledge in one place. You can find this information scattered throughout the amazing work of the researchers and books from many of the industry’s brightest minds, but never in one place.

In this chapter I will walk you through the process of taking your proof of concept into production, by detailing the implementation and deployment phases of an RL project. By the end I hope that these ideas will resonate and you will have a broad enough knowledge to at least get started, and understand what you need to dig in further. With this chapter, and of course the whole book, my goal is to bring RL to industry and demonstrate that production-grade industrial RL is not only possible, but lucrative.

In the first half I review the implementation phase of an RL project, looking deeper into the available frameworks, the abstractions, and how to scale. How to evaluate agents is also prominent here, because it is not acceptable to rely on ad hoc or statistically unsound measures of performance.

The second half is about operational RL and deploying your agent into production. ...