Enhancing Mobile Multitasking with Reinforcement Learning-Based Attention Management

Attention Management Systems promise to support users in better handling demands and mitigate the negative effects of multitasking. However, significant challenges remain in transferring these concepts to realistic settings. We introduce a reinforcement learning-based attention management system that schedules mobile operations with dynamic priorities, which we trained on a computationally rational simulation of typing behavior. To evaluate its effectiveness, we conducted two user studies, one in a stationary and one in a mobile context. Third, we conducted a simulation-based analysis comparing our RL-based policy with a heuristic baseline. The analysis demonstrated improved performance on several metrics for the RL-based policy as compared to self-supervision, including the number of words typed per minute and reaction times. The approach can successfully handle multiple concurrent tasks and worked in both the sitting and walking scenarios. We conclude by discussing implications for the design of Attention Management Systems.