To improve the efficiency of optimization via simulation, a new method called the generalized regression neural network based optimization via simulation (GRNN-OvS) is put forward. This method takes advantage of GRNN's non-linear approximation, learning speed and network stability, which is promising for predicting the simulation output if the neural network is fully trained by samples obtained from simulations. By means of substituting GRNN prediction for simulation output, the time spent in simulation via optimization can be greatly reduced. To be detailed, a certain amount of representative samples were initially generated from the simulation model. Then, the GRNN was trained using these samples so that the GRNN model can form a regression surface that provides good approximation of the input-output relationship of the simulation, which is considered as black box. Based on the GRNN model, the optimization via simulation problem is transformed to optimizing the GRNN model, which is more efficient and time-saving. Numerical example shows that GRNN-OvS is effective and feasible, which is very helpful in optimization via simulation applications where the simulation task is time-consuming.