Talk: On non-ergodicity in reinforcement learning

In reinforcement learning one considers the expectation value of future rewards in order to define an objective function for learning. The underlying probability measure of this expectation value is often coupled with the dynamics through a stationarity condition. In the non-ergodic case, however, this coupling does not uniquely determine the probability measure so that the objective function is not necessarily well-defined. My presentation will be mainly pointing to a number of questions that I find interesting.