Existing optimization methods for Generative Adversarial Networks that use evolutionary algorithms frequently suffer from the problem of disordered ordering, which occurs when several sets of solutions produced by a generator during training have poor ordering. To address the aforementioned issue, we want to optimize the generation process by training a Pareto-wise end-to-end ordered generator and ranking its children via evolutionary algorithms. Finally, the MCE-GAN algorithm is proposed to decompose the problem of adversarial optimization of generative adversarial networks into two problems of generator and discriminator evolution, and to perform mutation operations on the generator and discriminator using evolutionary algorithms, where the generator selection method is performed by training a set of generated generators and then using non-dominated ordering. Extensive studies on a synthetic dataset and a benchmark picture dataset demonstrate that the proposed MCE-GAN produces competitive and superior results.