Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Generating Search and Pointing Behavior on Grid Menus via Reinforcement Learning
The goal of this thesis is to employ state-of-the-art reinforcement learning techniques for user modelling in a human-computer interaction context. Specifically, we want to model how people search and operate grid menus.
Keywords: HCI, Reinforcement Learning, User Modelling, Computational Interaction
Evaluating an interface is expensive and hard, which normally requires many users. These users need to operate in a controlled setting, learn the interface and then use it on a day-to-day basis for proper evaluation. This is time-consuming, especially if the menu design is part of an iterative process. In an ideal world, a menu is optimized (cf. MenuOptimizer [1]), however a cost function is needed for this. Ideally, a cost function incorporates human behavior. We want to investigate whether Reinforcement Learning (RL) techniques are 1) suitable to approach human-like behavior and 2) can be part of an optimization process for optimizing menu layouts.
We want to explore a method that can search and operate a grid-style menu in a human-like fashion (an example of a grid-style menu is Windows 8). Current approaches do not take the full human into account and are not generative. This limits their use and interpretability in real-world applications. We plan to integrate existing closed-form models for perception (e.g. [2]) and action (e.g. Fitts law [3] ) to limit the action and observation spaces. To map from observation to action, we intend to use RL. This combination should enforce human-like, yet optimal (with regard to a reward function), behavior from the agent.
Work that will follow this thesis (or could be a stretch goal) will use this method of learning a human user model to optimize a grid menu. Therefore, it is important that the learned model can generalize to different menus, which will be a major challenge of this thesis.
Applicants should have some background knowledge in RL and an affinity or interest in computational HCI.
[1] Bailly, Gilles, et al. "Menuoptimizer: Interactive optimization of menu systems." Proceedings of the 26th annual ACM symposium on User interface software and technology. 2013.
[2] Acharya, Aditya, et al. "Human Visual Search as a Deep Reinforcement Learning Solution to a POMDP." CogSci. 2017.
[3] MacKenzie, I. Scott. "Fitts' law as a research and design tool in human-computer interaction." Human-computer interaction 7.1 (1992): 91-139.
Evaluating an interface is expensive and hard, which normally requires many users. These users need to operate in a controlled setting, learn the interface and then use it on a day-to-day basis for proper evaluation. This is time-consuming, especially if the menu design is part of an iterative process. In an ideal world, a menu is optimized (cf. MenuOptimizer [1]), however a cost function is needed for this. Ideally, a cost function incorporates human behavior. We want to investigate whether Reinforcement Learning (RL) techniques are 1) suitable to approach human-like behavior and 2) can be part of an optimization process for optimizing menu layouts.
We want to explore a method that can search and operate a grid-style menu in a human-like fashion (an example of a grid-style menu is Windows 8). Current approaches do not take the full human into account and are not generative. This limits their use and interpretability in real-world applications. We plan to integrate existing closed-form models for perception (e.g. [2]) and action (e.g. Fitts law [3] ) to limit the action and observation spaces. To map from observation to action, we intend to use RL. This combination should enforce human-like, yet optimal (with regard to a reward function), behavior from the agent.
Work that will follow this thesis (or could be a stretch goal) will use this method of learning a human user model to optimize a grid menu. Therefore, it is important that the learned model can generalize to different menus, which will be a major challenge of this thesis.
Applicants should have some background knowledge in RL and an affinity or interest in computational HCI.
[1] Bailly, Gilles, et al. "Menuoptimizer: Interactive optimization of menu systems." Proceedings of the 26th annual ACM symposium on User interface software and technology. 2013. [2] Acharya, Aditya, et al. "Human Visual Search as a Deep Reinforcement Learning Solution to a POMDP." CogSci. 2017. [3] MacKenzie, I. Scott. "Fitts' law as a research and design tool in human-computer interaction." Human-computer interaction 7.1 (1992): 91-139.
Create a suitable OpenAI Gym Environment, integrating knowledge from existing literature on biomechanical and cognitive constraints.
Employ RL to learn search, pointing and hand-eye coordination.
Compare to real users in a small user-study.
Stretch: Use the method to optimize an interface
Create a suitable OpenAI Gym Environment, integrating knowledge from existing literature on biomechanical and cognitive constraints. Employ RL to learn search, pointing and hand-eye coordination. Compare to real users in a small user-study. Stretch: Use the method to optimize an interface
Thomas Langerak - thomas.langerak@inf.ethz.ch
Sammy Christen - sammy.christen@inf.ethz.ch