Study on Stages Evaluation for 2D Action Game Generated by DC-GAN Based on Artificial Intelligence

Abstract

Recently, research on generative adversarial networks (GANs) in deep learning has advanced rapidly. For instance, in the field of image recognition, using a GAN, the number of training data was increased. GAN have also been used to create new similar images using training images, which can help designers such as car designer, character designer, game designer and more. We have been researching a method to automatically generate a game stage based on a dot-picture using GAN. So far, we have been studying game stage generation using GAN. In this study, we consider an evaluation method for generated game stages. Specifically, we consider using the A * search algorithm to evaluate the playability of the generated game stage from the number of jumps and the clear time. As a result, the jumping count of an automatically player using the A * algorithm was 17 times for the original stage, 12 times for stage No.1, 16 times for stage No.2, and 18 times for stage No.3. Next, the clear time was 9 seconds for the original stage, 8 seconds for stage No.1, 10 seconds for stage No.2, and 11 seconds for stage No.3. In other words, we suggest stage No.1 is simpler than the original stage, and stage 2 and stage 3 are a little more difficult than the original stage.