I have recently become a bit familiar with the machine learning techniques, and examples of problems where they are ought to be applied. For example, we can try deriving models for the time series or train our machines to predict their dynamics. For example, in reinforcement learning we are trying to develop a good controller for an unknown system while learning from this system.
It seems that system identification tries to approach a similar problem: at least in most of the cases I heard of one assumes that a system in question (e.g. wind power generator) is a linear dynamic system with some stochastic input (wind in this case) and the task is to estimate the transfer function - e.g. in discrete time the coefficients and the depth of the ARMA model.
Since these two fields have such similar methods and problems, I would expect them to be very well connected, however it seems that two communities are rather separated. Is that true? In such case, what are the major differences in their approaches? A related question was asked on TCS, but have not received much attention there (unsurprisingly as it's not that close to theoretical compute science).