The simulation of user behavior with deep reinforcement learning agents has shown some recent success. However, the inverse problem, that is, inferring the free parameters of the simulator from observed user behaviors, remains challenging to solve. This is because the optimization of the new action policy of the simulated agent, which is required whenever the model parameters change, is computationally impractical. In this study, we introduce a network modulation technique that can obtain a generalized policy that immediately adapts to the given model parameters. Further, we demonstrate that the proposed technique improves the efficiency of user simulator-based inference by eliminating the need to obtain an action policy for novel model parameters. We validated our approach using the latest user simulator for point-and-click behavior. Consequently, we succeeded in inferring the user’s cognitive parameters and intrinsic reward settings with less than 1/1000 computational power to those of existing methods.

We present a novel simulation model of point-and-click behaviour that is applicable both when a target is stationary or moving. To enable more realistic simulation than existing models, the model proposed in this study takes into account key features of the user and the external environment, such as intermittent motor control, click decision-making, visual perception, upper limb kinematics and the effect of input device. The simulated user’s point-and-click behaviour is formulated as a Markov decision process (MDP), and the user’s policy of action is optimised through deep reinforcement learning. As a result, our model successfully and accurately reproduced the trial completion time, distribution of click endpoints, and cursor trajectories of real users. Through an ablation study, we showed how the simulation results change when the model’s sub-modules are individually removed. The implemented model and dataset are publicly available.

In virtual reality (VR), a user's virtual avatar can interact with a virtual object by colliding with it. If collision responses do not occur in the direction that the user expects, the user experiences degradation of accuracy and precision in applications such as VR sports games. In determining the response of a virtual collision, existing physics engines have not considered the direction in which the user perceived and estimated the collision. Based on the cue integration theory, this study presents a statistical model explaining how users estimate the direction of a virtual collision from their body's orientation and velocity vectors. The accuracy and precision of virtual collisions can be improved by 8.77% and 30.29%, respectively, by setting the virtual collision response in the direction that users perceive.