TFTune: Creation and Personalization of Pointing Transfer Functions Using Reinforcement Learning

Paper Title

TFTune: Creation and Personalization of Pointing Transfer Functions Using Reinforcement Learning

Publication Info

• Topic area: Reinforcement learning-based optimization of pointing transfer functions for input devices.
• Keywords: Transfer functions, reinforcement learning, personalization, pointing devices, human-computer interaction, macOS, Windows, muscle-computer interface, performance optimization, user studies.

Background and Problem

• Problem / challenge: Existing pointing transfer functions have seen limited improvements since the introduction of acceleration-based gains. Current systems offer minimal personalization, and prior automated tuning approaches fail to outperform default operating system functions.
• Significance: Pointing is a fundamental interaction for millions of users daily. Improving transfer functions can enhance productivity and usability, especially for diverse devices and user needs.
• Motivation and related work: Previous work on transfer functions has focused on static or manually tuned mappings, with limited success in improving performance. Reinforcement learning (RL) has shown promise in human-in-the-loop optimization but has not been applied to transfer function tuning. This paper addresses the gap by introducing a scalable, RL-based solution.

Solution

• Proposed approach: TFTune, a reinforcement learning-based system for creating and personalizing pointing transfer functions, optimizing performance for various input devices and user contexts.
• Novelty:
  1. First RL-based approach to significantly improve pointing transfer functions beyond OS defaults.
  2. Capable of creating transfer functions from scratch and personalizing existing ones.
  3. Generalizes across diverse hardware and input modalities, including muscle-computer interfaces (MCIs).
  4. Achieves rapid convergence (<10 minutes) with minimal user input.
• Procedure and key techniques:
  • Models transfer function tuning as a Markov Decision Process (MDP).
  • Uses Proximal Policy Optimization (PPO) to train a neural network policy that maps input states to gain values.
  • Converts the trained policy into a deployable transfer function via lookup tables or direct inference.
  • Evaluates performance through user studies on macOS, Windows, and MCIs.

Results

• Concrete findings:
  • TFTune improved movement times by 7% on macOS (trackpad), 8% on Windows (mouse), and 16% for MCIs.
  • Tuning times were approximately 7 minutes for macOS, 1 minute for Windows, and 2 minutes for MCIs.
  • Participants reported no performance degradation after 2–4 weeks of at-home use.
• Advantage over baselines:
  • Outperformed macOS defaults by 7% and AutoGain by 5% in Study 1.
  • Outperformed Windows defaults by 8% in Study 2.
  • Reduced movement time by 16% compared to proportional control in MCIs during Study 3.
• Experiments / evaluation:
  • Study 1: Compared TFTune to macOS defaults and AutoGain on a MacBook trackpad with 12 participants.
  • Study 2: Tested TFTune on participants' personal Windows devices with 14 participants, including a 2–4 week at-home deployment.
  • Study 3: Evaluated TFTune on a rate-controlled MCI with 10 participants using an EMG-based interface.
• Limitations and future work:
  • Slightly higher error rates observed in some conditions, potentially due to reward function design.
  • Did not explore how personalized functions might inform global transfer function improvements.
  • Future work includes refining reward functions, enabling continuous tuning during real-world use, and exploring richer input features.

Summary

TFTune is a reinforcement learning-based system for creating and personalizing pointing transfer functions, achieving significant performance improvements over default OS functions and prior approaches. It demonstrated scalability across diverse hardware and generalizability to new input modalities like muscle-computer interfaces. User studies showed rapid convergence and sustained performance gains, with participants reporting no degradation during extended at-home use. This work highlights the potential of RL for optimizing fundamental human-computer interactions and opens avenues for further research in transfer function design and personalization.