We're using it for web crawling: define what to look for (a reward function), and crawler can learn how to get these pages from the web without wasting too much HTTP requests for irrelevant content. No neural nets, just Q-Learning with linear function approximation, with some common tricks like double learning and experience replay.
I work on a few algorithms that could be classified as RL given an open mind. Most of them learn distributions from streaming data via some kind of online EM. I know that people in the ad-serving, porn-serving, and website optimization (A/B stuff) sectors use RL pretty extensively as well, but I'm not one of them at the moment.
> learn distributions from streaming data
That's unsupervised learning afaik - clustering, manifolds etc. Where is "reinforcement" part there (agent, environment, reward)?
If you use RL, you might not know it. Multi-armed bandits, for example.
Used simple RL (finite states, no neural nets) for gamedev - finding optimal policies for player behavior helps in fixing balance issues.
+1, I also am curious of how prevalent is RL in industry.