I wonder, too. My understanding of RSM is that it is a methodology for the minimization of an unknown function which however can be evaluated at given points in space - typically through a computer simulation of some system. Given that one has evaluated a few such points (i.e., having selected some vectors in space and run those simulations) one forms an explicit function that maps - as well as possible, based on some suitable metric - the evaluated points. The next step is to minimize this explicit function, which then yields a new point in space that can be added to the former ones in providing an updated (improved) explicit approximation of the function. Termination criteria depend on the properties of the function, I suppose.
RSM is a collection of mathematical methods to relate a response (dependent) variable to the independent ones in an experiment. So as a result, RSM is strictly related to DOE. The mathematical model used for the approximation of experimental data in the experimental domain is called "response surface" or "meta model". This method takes the advantage of representing a model to estimate the data over the design space as well as optimization according to the desired criteria. These definitions were presented according to the book you can find it through the link below.