Deep Learning for PCG of 2D video game levels

Content generation is one of the most expensive and time consuming tasks in video game development, and Procedural Content Generation (PCG) approaches have been proposed to reduce costs since 1980 in video games as Elite. In recent years, Generative Adversarial Network (GANs) have shown promising results in providing new possibilities for procedural generation of new content, including generation of 2D game levels. However, GAN research for PCG is still an open issue, since the new levels procedurally generated should respect constraints of the gameplay rules and they should be playable. In this paper we will review the state of the art and explore three interesting approaches in literature for procedural generation of game levels with GANs. First, we examine how the data of human-designed video game levels can be represented inside data-sets suitable for GANs; in particular, we focus on the representation of the game rules associated with each object present in the level. Second, we expose three different approaches of PCG with GANs, and for each approach we deepen the proposed adaptations on the GANs to generate video games levels, the data-set used, and the algorithm’s performance. For each approach, we also analyze the proposed evaluation methods for the procedural generated levels, in particular taking into account the traversability, verifying that the areas in the level are reachable, and the playability, verifying that the goal target of a game level is reachable without excessive difficulty.