Evolving Levels for General Games Using Deep Convolutional Generative Adversarial Networks

Abstract

Deep Convolutional Generative Adversarial Networks (DCGANs) are a machine learning approach that can learn to mimic any distribution of data. DCGANs consist of a generator and discriminator where generator generates new content and discriminator finds whether the generated content is real or fake. Procedural Content Generation (PCG) for level generation could potentially benefit from such models, particularly for a game where there is an existing level to emulate. In this paper, DCGANs are trained on existing levels generated through a random generator. Three different games (Freeway, Zelda, and Colourescape) are selected from the General Video Game Artificial Intelligence (GVG-AI) framework for level generation. The proposed approach successfully generates various levels which mimic the training levels. Generated levels are also evaluated using agent-based testing to ensure their play-ability.