AI-Driven Analysis and Optimization of Fairness in Competitive Video Games

Since their inception, video games have experienced an immense growth in popularity. A once-niche hobby a few partook in is now one of the most accessible and widespread forms of recreation. The market has adapted to such demand: in 2023, the video game market size was estimated at over 240 billion USD and projected to reach 650 billion within 10 years. Video games can be costly to develop, and the titles with the highest budgets, known as AAA, can reach comprehensive costs ranging in the hundreds of millions USD.

Much like any piece of software, video games require testing. Due to the complexity of modern video games, a few tests by the developer do not suffice, and entire teams are specialized in quality assessment (QA). QA of a video game can account for as much as 20% of the total production costs.

One of the most important aspects of QA is improving game balance, a factor crucial for the success of a game. Especially for competitive titles, unfair advantages can easily lead to frustration and, consequently, low player retention.

In this project, we experiment with training AI agents tasked with evaluating the fairness of a game as a bounded metric. We then proceed with an automatic fairness-driven tuning of the game’s configuration to find the configuration that achieves maximum fairness. This automation could potentially reduce the reliance on human QA, thus optimizing development costs and time.

To validate our approach, we review and compare the data and results obtained from a human player population against those of AI agents, checking for similarities. We carry out this experiment on the Pokémon Showdown platform due to the tractability of the game and active player base.