Reinforcement learning in business

As our team progresses on our first integrated product, and having gone through an extended period of reading, studying, and corresponding with Microsoft Research and DeepMind, we've also begun numerical experiments probing how full reinforcement learning tools could help businesses in spaces like sign-up flows and chatbots. But as we recreate the work of the most famous papers in this field, something strange keeps happening: the claims of these papers just don't hold up outside of their toy problems.

I mean, I shouldn't be too surprised. The promise of deep learning has been way overblown, and many business applications are outperformed by easier-to-manage algorithms like gradient boosted trees. But since the academic work in reinforcement learning limits itself to unrealistic toy problems, and the corporate papers follow suit, finding more and more complicated algorithms and proofs for the same problem space, they keep avoiding the questions any actual customer would come up with.

It's this middle ground -- between the simplicity of a bandit service, and the complexity of a chess algorithm, that is severely lacking in the literature. By building our tools as customers first, we hope to crack this problem and give the world something to really cheer for. Stay tuned!