Machine-Learned Locomotion and its Effect on Player Experience in Games

Sammanfattning

This thesis investigates how machine-learnedlocomotion affects player perception of movementbehaviour and gameplay experience in game agents.To investigate this, a comparative case study wasconducted between two quadrupedal agents developedin the Unity game engine. One agent utilizedreinforcement learning to produce a physics-basedmovement behaviour while its counterpart utilizedtraditional keyframed animations combined with afinite state machine with animation blending. Bothagents were evaluated in a interactive gameplayscenario where participants were required to completegameplay objectives while under pursuit of one of theagents. This was meant to reflect a typical Playerversus environment (PVE) experience. The study combines quantitative questionnaire datawith qualitative participant responses in order toevaluate how machine-learned locomotion inperceived by players and what effects it has on theplayer experience. To evaluate this, player perceptionof agent locomotion was analyzed across severalcategories, including perceived realism, intentionality,responsiveness, naturalness, intelligence,predictability, threat, and overall enjoyment. The findings suggest that machine-learned locomotionand traditional animation systems provide differentstrengths from a player experience perspective. Themachine-learned agent was frequently perceived asmore dynamic, natural, unpredictable and engaging tointeract with during gameplay. In contrast thetraditionally animated agent was perceived as moreresponsive, predictable, intentional and its movementwas perceived as clearer to players. Despite themachine-learned agent displaying instability and inconsistent movement behaviour at times, a majorityof players reported preferring to play against themachine-learned agent often due to its dynamic andless repetitive movement behaviour.The study contributes to prior research by shiftingfocus from technical aspects and improvement toreinforcement learning algorithms and toward playerperception of such implemented systems in agameplay environment, in order to gain insight onhow such a model is perceived during interactivegameplay.