An Automatic Approach to Transfer Function Tuning Using Reinforcement Learning

Optimizing an interaction to achieve the highest possible performance is desirable for many applications. For instance, even minor improvements in the mapping between a computer mouse’s movement and a cursor on a screen can significantly enhance comfort and usability. Similarly, better mapping an operator’s joystick inputs to crane movements could boost productivity. Both examples emphasize the importance of finding an ideal transfer function (i.e., the mapping between human input and system output). However, while we have the affordance of relying on several decades of tuning when using the transfer function of touchpads and mice, we do not have this convenience when it comes to novel interaction techniques. Correspondingly, we need approaches to automatically determine the ideal transfer function to create the most performant interactive systems possible. We propose using reinforcement learning (RL) to identify optimal input-output mappings for novel inputs. Our approach will first explore familiar devices like touchpads and mice before extending to emerging inputs like electromyography. This work will be supported by a collaboration between a French research lab specializing in transfer functions and a Canadian lab focused on machine learning. Together, we aim to advance the automation of transfer function tuning, creating more efficient and accessible systems.

Faculty Supervisor:

Scott Bateman

Student:

Partner:

Université de Lille

Discipline:

Engineering

Sector:

Artificial Intelligence

University:

University of New Brunswick

Program: