Simplex gradient algorithm has the shortages of slow computational speed and strong randomness of well controls obtained by algorithm in reservoir injection-production optimization, etc. To solve these problems, a modified simplex gradient algorithm is therefore proposed. By choosing appropriate perturbation variables, the modified simplex gradient is approximately equal to the product of well control covariance matrix and the true gradient. Thus, the correlation between the changes of well controls and the control time steps is considered. In addition, the preconditioned strategy is applied in solving the gradient, which avoids the singular value decomposition and pseudo-inverse computation for large matrix. By adjusting the operating production conditions of oil and water wells in each control steps of reservoir production dynamically, the computational efficiency has been significantly improved, and the control result obtained is smoother, which can be operated in practical production applications easily. Moreover, the economic benefit optimized by modified algorithm increases about 10% more than original simplex gradient algorithm. The results verify the effectiveness and feasibility of the algorithm.