Embedded Files
**Update: Tutorial video is now uploaded on YouTube: https://www.youtube.com/watch?v=bHBH2qpINoQ
Tutorial Overview
Reinforcement learning (RL) addresses sequential decision-making problems. In RL an agent interacts with its environment in a trial-and-error fashion to learn an optimal policy for a given task. Despite the remarkable success achieved by RL in various domains, the adoption of RL in real-world environments is still very limited due to both technical and social challenges. Technically, the need for an RL agent to excessively interact with the environment to learn an effective policy is a significant challenge due to the associated time and resource consumption. Socially, the limited ability of RL agents to cooperate with their human teammates and to ensure socially acceptable behaviours hinders the acceptance of RL by its potential users.
Designing techniques for users and domain experts to intuitively teach RL agents is an effective means for addressing these challenges. Such teaching techniques enable the direct transfer of domain knowledge which leads to efficient learning and consequently facilitates the acceptance of the resulting RL agents’ behaviours.
This two-hour tutorial aims to provide an introduction to human teachable RL. The audience will be :
introduced to RL and the key challenges that hamper its adoption
exposed to the existing techniques for teaching RL agents by lay users and domain experts to address these challenges, and
outlined with the open questions and promising research directions in this area.
Who should attend?
The target audience of this tutorial are researchers and practitioners interested in reinforcement learning and/or human-machine interaction. Previous experience with reinforcement learning will be helpful, but not necessary, for attendees to make the best of the tutorial. The first component of the tutorial is designed to allow attendees with little-to-no reinforcement learning background to follow up and understand most of the concepts covered by the tutorial.
Description & Outline
The tutorial is divided into three parts as describe below:
Part 1: Introduction to Reinforcement Learning [30 minutes]
This component aims to provide a background for audience with little-to-no experience in RL. It will then present key challenges of adopting RL in real-world settings. This component will cover:
a. what RL is and how it is different from other machine learning approaches
b. RL as a Markov Decision Process
c. applications & key challenges of RL
Part 2: Human-Teachable Reinforcement Learning
This component will present an overview of existing techniques for teaching RL by lay users and/or domain experts. This component will cover:
a. the need for human teachable RL and
b. existing techniques for human teaching to RL agents
Part 3: Towards Comprehensive Teaching Approaches for Reinforcement Learning Agents
The last component provides an overview of the limitations of the current RL teaching technique and Identifies promising future extensions. This component will cover:
a. limitations of the existing teaching techniques in relation to the sophistication of recent RL agents
b. promising future directions & open research questions
Presenter
This tutorial will be presented by Dr. Aya Hussein - Research Associate, School of Engineering & IT, University of New South Wales, Canberra, Australia
Biography
Aya Hussein is a research associate at the School of Engineering and IT, University of New South Wales, Canberra, Australia. Her research focuses on combining human and machine intelligence by studying interaction schemes that allow humans to teach machines and to effectively team up with them in the field. Particularly, Aya is interested in researching techniques for teaching reinforcement learning agents by humans rather than by learning completely on their own via trial and error. Aya received her PhD in Computer Science from UNSW- Canberra in 2020. Her PhD was about human-swarm teaming where she studied both human and machine factors that impact team effectiveness. Prior to that, Aya got her master’s and bachelor’s degrees in Computer Engineering from Cairo University in 2011 and 2015.
Expertise in the Tutorial Area
Aya Hussein has solid expertise in the fields of human-teachable Reinforcement Learning and Human-Machine Interaction. She has 19 publications in these and related areas. She has co-organised a workshop on “Cognitive and Social Aspects of Human Multi-Robot Interaction” in the flagship conference of IEEE IROS 2021. She has reviewed 40 articles for many journals and conferences including IEEE Transactions on Neural Network and Learning Systems, IEEE Transactions on Cognitive and Developmental Systems, and IEEE Transactions on Artificial Intelligence.
Aya has been teaching an undergraduate course on Reinforcement Learning in 2022. She has been supervising 11 undergraduate and post-graduate research projects in related topics.
Page updated
Google Sites
Report abuse