Modelling Human Behaviour for Believable NPCs

Abstract

This thesis explores modelling believability for non-player characters (NPC) when fulfilling the role typically reserved for human players in the first-person shooter (FPS) genre. Numerous experiments are conducted, which are focused on acquiring specific data on gameplay patterns and behaviours that emerge when undertaking tasks and actions in a deathmatch scenario. The original contribution of this thesis is centred on the approach of using player-driven data and traditional development techniques to develop NPCs indistinguishable from human players.

NPCs have become an essential aspect of a video game; they are used to create an immersive and vibrant environment, and depending on their purpose, can significantly impact how the game is perceived. When NPCs fulfil a role typically reserved for a human player, if they do not imitate a player correctly, they can have a detrimental effect on the experience. In team-based games with consequences to be eliminated, they can be a negative to their team.

The research undertaken in this thesis breaks down gameplay into core aspects so that they can be individually explored and discussed. The focus is on determining the extent of emerging gameplay patterns if these patterns can be modelled, and if so, how well these models function in a deathmatch scenario. Perception is a crucial feature of this project because believability can only be maintained if NPCs can visually and intuitively pass as human players. When believability collapses, it is difficult to regain it.

The research questions proposed in this thesis are answered through experiments and questionnaires. It shows that gameplay can be generalised, and to an extent, patterns emerge. The resulting models indicate an improvement over non-modelled NPCs, with NPCs appearing more human-like with each iteration. A series of perception tests show that human players are aware of their surroundings and opponents, and can intuitively know when something is not as it should be. Furthermore, poorly implemented NPCs can impact the overall enjoyment.

Date of Award	2025
Original language	English
Supervisor	Christopher Tubb (Supervisor), Andrew Ware (Supervisor) & Emlyn Everitt (Supervisor)