In this study, we want to develop an efficient optimization algorithm based on Perturbation Analysis (PA) and compare the performance of PA with the traditional Finite Difference (FD) gradient method. While the FD estimate is a very time‐consuming gradient method, PA is a very promising algorithm in terms of the computational time. This is the major advantage of PA over FD. In this study, we use the property that PA only requires one simulation run to obtain other advantages of PA as a gradient method. First, we deal with the problem of initialization bias for the estimate of gradients of system performance. Second, we compare the efficiency of PA and FFD (which represents FD) in introducing the control variates variance reduction technique. Third, we propose a new method for determining the simulation run length at each iteration in an optimization algorithm. In the above three areas, we show that PA is better than FD. Finally, we combine PA and BM (batch means method) to obtain a more efficient gradient method, PA‐BM.