Instant Difficulty Adjustment using User Skill Model Based on GPDM in VR Kendama Task

Abstract

Adapting to user’s skill is crucial to task difficulty adjustment. This paper presents a task difficulty adjustment method that predicts future success rate when changing the difficulty level with small data from each user: instant difficulty adjustment. We proposed a methodology based on a Gaussian process dynamical model (GPDM) to model the user’s skill from past performance observations, and predict future performance at a targeted difficulty level stochastically. As a task to be performed, we focused on a cup-and-ball game (a kind of juggling called Kendama) using virtual reality (VR), in which the cup size is changeable to adjust the difficulty level in a VR environment. In the experiment, we compared the proposed method with LSTM-based deterministic method set by randomized initial parameters with participants who had different skills of the Kendama task. Our results indicate that the GPDM-based method accurately reflects the user’s skill, and the predicted success rate is close to the actual success rate, which tends to appear under the conditions of balanced training data on the number of successes and failures. Additionally, our method is valid for decreasing the training data, which means the prediction accuracy is ensured even with a small number of Kendama trials. In future work, we will achieve the instant difficulty adjustment at various training data not restricted to the number of successes and failures.