Many approaches to solving optimization problems with zeroth-order oracle are constrained by non-smooth object functions. In this work we are focused on smooth problems with only zero-order information accessible. Such tasks are still viable, since sometimes we can not use autograd for obtaining first order information (art.1). The approach is based on "stohastic smoothing" (art.2) using stohastic gradient for gradient approximation. This article provides iteration and zeroth-order oracle call number bounds.
All work and obtained results are located in Posted.pdf