From Predict to Control & From RL to Offline RL


link to the PPT

In this talk, I use recommender systems as a starting point to build a bridge between the prediction problem and the decision-making problem (a.k.a. optimal control). Then I present the general idea of reinforcement learning and some recent progresses. Lastly, I introduce offline reinforcement learning, which is an important direction towards real-world reinforcement learning, aiming to avoid the expensive cost on online learning and copy the success of DL+Big Data.