Grating stereoscopic display is currently one of the most popular naked eye three-dimensional(3D) displays, and optimizing its parameters is a key step in the display design process. This article proposes an optimization method for grating stereoscopic displays based on simulation image evaluation. This method is based on existing grating stereoscopic display simulation platforms and assesses display quality by evaluating simulation images, which is more intuitive and convenient, rather than optical indicators. This research invites reviewers to conduct subjective experiments and selects the most suitable evaluation criteria for 3D display simulation images: learned perceptual image patch similarity (LPIPS). A joint optimization experiment is conducted on the parameters of a cylindrical lens grating stereoscopic display using Bayesian method, and compares with the traditional optical optimization results of Zemax. The diffuse spot radius of the center field of view is 15.567 μm, and the diffuse spot radius of the edge field of view is 89.744 μm, both of which meets the design standards of empirical requirements. The experiment results show that this method can correctly optimize the commonly used parameters of the display and achieve the same effect as traditional optical design software. This method has significant potential in the optimization design of grating stereoscopic displays, further provides a new approach for the design process of naked eye 3D displays.