Game Balancing via Procedural Content Generation and Simulations

Abstract

Balancing games requires extensive manual work and human playtesting during development. Existing research proposes using search-based optimization combined with game simulations to estimate balance. However, these approaches are tailored to specific environments, making them difficult to transfer to other games. Additionally, simulating a game is computationally intensive. My research aims to develop procedural content generation (PCG) methods to automate the generation of balanced content for competitive games. Unlike related work, domain-independent methods enable transferability to other games. I focus on two key aspects that significantly impact a game's overall balance: game levels and economies. Game level balancing is approached by framing it as both, a PCG task and a Markov decision process in order to apply reinforcement learning (RL). Therefore, I adapt and extend the PCGRL (PCG via RL) framework. Based on an existing formal definition, my research considers game economies as graph structures. In this context, I present two frameworks: G-PCGRL (Graph-PCGRL) and GEEvo (Game Economy Evolution).