A Game Balance Optimization Framework using Reinforcement Learning and Genetic Algorithms

Andersen, Patrick Nicolai

4. term

Medialogy, Master

2023

This thesis addresses the complex and time-consuming task of balancing games during development. It investigates whether a framework that combines reinforcement learning to train game-playing agents with a genetic algorithm to search the space of balance parameters can support designers in reaching fair and enjoyable game states. The framework incorporates user and designer input to preserve creative control and narrow the search space. To examine the approach, a game was developed and used to apply the methodology, with balance defined as achieving an equal win rate between two factions. The work presents a methodology that can be integrated into the game development cycle to simulate and evaluate balance proposals, while also noting several shortcomings and caveats in its application.

Dette speciale adresserer den komplekse og tidskrævende opgave med at balancere computerspil under udvikling. Formålet er at undersøge, om en ramme, der kombinerer forstærkningslæring til at træne spilagenter med en genetisk algoritme til at søge i spillets balanceparametre, kan støtte designere i at finde mere fair og sjove spiltilstande. Rammen integrerer også bruger- og designerinput for at bevare kreativ kontrol og indsnævre søgeområdet. For at udforske tilgangen er der udviklet et spil, hvor målet for balancen er defineret som en lige stor sejrsrate mellem to fraktioner. Specialet præsenterer en metode, der kan indgå i spiludviklingsprocessen og bruges til at simulere og evaluere forskellige balanceforslag, men det fremhæver også flere begrænsninger og forbehold i anvendelsen.

[This apstract has been generated with the help of AI directly from the project full text]

machine learning ; reinforcement learning ; game design ; game balancing ; genetic algorithms ; game development

Download PDF
View record in AAU Student Projects