When minimizing stochastic functions without reliable derivatives, many algorithms 1) repeatedly sample the function at a point of interest to more accurately determine the function value, 2) define a fixed pattern of points where the function will be evaluated within regions of interest or, 3) require the user to define a decaying sequence of step sizes to be used by the algorithm. In this talk, we will explain why we consider these three properties to be undesirable in a practical algorithm. We develop a trust region method which does not repeatedly sample points, does not require a predefined pattern of points, and adjusts the trust region radius as the algorithm progresses. After analyzing our method's stochastic convergence properties, we show it to be competitive with existing algorithms on a range of problems.