Assessing Multiplayer Level Design Using Deep Learning Techniques

Abstract

This paper proposes a new framework to measure the fairness of asymmetric level-design in multiplayer games. This work achieves real time prediction of the degree to which asymmetric levels are balanced using deep learning. The proposed framework provides both cost and time savings, by removing the requirement of numerous designed levels and the need to gather player data samples. This advancement with the field is possible through the combination of deep reinforcement learning (made accessible to developers with Unity's ML-Agents framework), and Procedural Content Generation (PCG). The result of this merger is the acquisition of accelerated training data, which is established using parallel simulations. This paper showcases the proposed approach on a simple two player top-down-shooter game implemented using MoreMountains: Top Down Engine an extension to Unity 3D a popular game engine. Levels are generated using the same PCG approaches found in 'Nuclear Throne' a popular cross platform Roguelike published by Vlambeer. This approach is accessible and easy to implement allowing games developers to test human-designed content in real time using the predictions. This research is open source and available on Github: https://github.com/Taikatou/top-down-shooter.