ADAPTABLE ADVERSARY BEHAVIOUR FOR GAME OF TAG BASED ON PLAYER BEHAVIOUR USING MACHINE LEARNING

Abstract

Adversaries in video games would seldom be created to rely on predefined behaviour trees to respond to various player behaviours and fulfil simple goals in order to deliver an experience designed by the video game developer. While this is adequate for most games, immersion and the intended experience of the video game would be compromised when the adversary is faced with unexpected player behaviour, which would usually end up confusing the adversary’s artificial intelligence (AI). This occurrence is known as a “sequence break”, which is an event that occurs when the behaviour tree is poorly designed and contains known exploits that could either significantly reduce the difficulty of dealing with the adversaries or allow the player to bypass them entirely. Allowing adversaries to adapt to player behaviour as the game progresses using adversarial reinforcement learning would provide a dynamically evolving experience in which the user would be continuously challenged depending on their cleverness, forcing them to play the game as intended without the use of exploits. Inadequate adversarial AI in video games can make a game more boring or frustrating depending on the set difficulty level determined by the game developer. In this study, past and current implementations of AI in video games are explored, including the unique innovations which can be found throughout the video game industry regarding video game AI adversaries.
Keywords: artificial intelligence, adversarial reinforcement learning, video games, behaviour tree