Content evaluation of StarCraft maps using Neuroevolution

Abstract: Context. Games are becoming larger and the amount of assets required is increasing. Game studios turn toward procedural generation to ease the load of asset creation. After the game is released the studios want to extend the longevity of their creation. One way of doing this is to open up the game for community created add-ons and assets or utilize some procedural content generation. Both community created assets and procedural generation comes with a classification problem to filter out the undesirable content. Objectives. This thesis will attempt to create a method to evaluate community-generated StarCraft maps with the help of machine learning. Methods. Manually extracted metrics from StarCraft maps and ratings from community repositories. This data is used to train neural networks using NeuroEvolution of Augmenting Topologies (NEAT). The method will be compared with Sequential Minimal Optimization (SMO) and ZeroR. Results and Conclusions. The problem turned out to be more difficult than initially thought. The results using NEAT are marginally better than SMO and ZeroR. The suspected reason for this is insufficient input data and/or bad input parameters. Further experimentation could be conducted with deep learning to try to find a suitable solution for this problem.