Improving Feedback in Human-in-the-loop Reinforcement Learning


75472308 036a 4f15 81be bf6791839a4b.png 500 500

Organization: Advanced Interactive Technologies

Involved Host(s): Christen Sammy

Abstract: In this project, we would like to explore whether an attention or saliency visualization technique can improve teacher feedback in a human-in-the-loop reinforcement learning scenario.

Description: Approaches that use human-feedback to guide RL-agents suffer from poor scalability. Furthermore, it is hard to make an informed feedback decision as a user about whether an agent took some actions for the right reason or not. Therefore, we want to study if a visual interpretation of an agent’s policy, in the form of an attention or saliency mechanism, can help users make better feedback decisions. Can a user distinguish a good policy from a bad policy when given a visualization mechanism? The main task is to find a suitable visualization technique and run a comparative study on RL-agent policies. In an extended step, the goal is to see whether such an augmentation of the observations can improve the performance of human-in-the-loop RL algorithms in game-based environments such as Atari.

Contact Details: Sammy Christen (sammy.christen@inf.ethz.ch) Dr. David Lindlbauer (david.lindlbauer@inf.ethz.ch)

Reinforcement Learning Computer Vision Saliency Attention Human-in-the-loop


Labels: Semester Project Bachelor Thesis Master Thesis CLS Student Project (MPG ETH CLS) ETH Organization's Labels (ETHZ)
Topics: Engineering and Technology Behavioural and Cognitive Sciences
Applicant Organizations: ETH Zurich Max Planck ETH Center for Learning Systems