How Reinforcement Learning can improve Video Games Development: Dreamer and P2E Algorithms in the SheepRL Framework

Artificial Intelligence (AI) in video games is along-standing research area. It studies how to use AI technologies to achieve human-level performance when playing games. For years now, ReinforcementLearning (RL) algorithms have outperformed the best human players in most video games. For this reason, it is interesting to investigate whether RL can still be used in the video game industry or whether the relationship between RL and the video game industry should remain purely academic. This work focuses on two primary objectives within the video game industry: (i) Testing and Debugging: how RL can be exploited in order to uncover latent bugs, assess game difficulty, and refine the design of the video game. (ii) Non-Playable Characters (NPC) Creation and Generalization: is RL the best strategy to efficiently create NPCs or the RL algorithms have become too advanced? This thesis explores the feasibility of using the state-of-the-art Dreamer algorithm in automated testing and NPCs creation for video games; in addition, it proposes SheepRL a scalable open source framework for running experiments in a distributed manner.