Details

Human-Like Behaviour

Year: 2025

Term: Winter

Student Name: Clint Galvez

Supervisor: Yuhong Guo

Abstract: This research project addresses the feasibility and challenges of creating human-like AI agents in video games; specifically focusing on developing agents that can follow targets while remaining hidden. Traditional game AI often becomes predictable, diminishing player immersion and enjoyment over time. To overcome this limitation, a 3D game with complex graphics and mechanics alongside an environment inspired by stealth-based games was developed. Featuring multiple navigation pathways with distinct risk-reward dynamics. The environment incorporates variable-height obstacles, strategic cover points, and open areas to encourage complex decision-making. Using Unreal Engine's Learning Agents plugin, reinforcement and imitation learning approaches were implemented with carefully designed reward systems that balance target visibility with stealth. The agent's perception system processes self-position, target location, and environmental data to inform decision-making, while the NPC employs sight, hearing, touch, and prediction capabilities to create challenging scenarios for the learning agent. The results demonstrate that the AI successfully learned behaviours that bridge the gap between rigid, scripted bots and complex human players. Eventually concluding that creating convincing human-like AI is most effectively achieved through a hybrid approach that synthesizes neural networks with behaviour trees, where machine learning focuses on selecting appropriate predefined behaviours rather than controlling low-level continuous actions. This modular system offers a pragmatic and scalable solution for developing immersive, adaptive AI in games.