Evaluating participant responses to a virtual reality experience using reinforcement learning.

Area of Science:

• Human-Computer Interaction
• Virtual Reality User Experience
• Experimental Psychology

Background:

• Virtual reality (VR) quality of experience (QoE) is influenced by rendering, responsiveness, and interface factors.
• Traditional post-exposure questionnaires have limitations, including framing effects and inability to capture non-additive factor interactions.
• Factorial experimental designs struggle with non-additivity for more than two factors.

Purpose of the Study:

• To develop and evaluate a novel reinforcement learning (RL) agent-based method for assessing VR user experience.
• To overcome limitations of traditional questionnaires and complex factorial designs in evaluating multi-factor VR QoE.
• To identify user preferences for specific VR application factors.

Main Methods:

• An RL agent was developed to dynamically propose changes to VR factor levels during user exposure.
• Participants interacted with the RL agent, accepting or rejecting proposed factor changes.
• An experiment with 20 participants evaluated four binary factors: navigation, avatar representation, virtual character responsiveness, and rendering style.

Main Results:

• The RL agent converged on a stable policy indicating user preferences.
• Participants consistently preferred teleportation for navigation over walking-in-place.
• Preferences included full-body avatars, responsive virtual characters, and realistic rendering over cartoon styles.

Conclusions:

• The RL agent-based method effectively evaluates user preferences for VR factors, capturing nuanced interactions.
• This approach offers a dynamic and less biased alternative to traditional methods for VR QoE assessment.
• The identified factor configuration provides insights for designing more engaging VR experiences.