URL: https://mcgillmrl.github.io/transfer_club/
MailingList: https://groups.google.com/forum/#!forum/transfer_club
Location: MC 627 McConnell Engineering Building
Time: Fridays at 4:30 p.m.
Our study group is concerned with the following two topics:
- Transfer Learning in robotics
- Hierachical Reinforcement Learning in robotics
We think these two problems are crucial forpractical applications of learning algorithms in robotics since collecting data with real robots is expensive and the search space for policies in real robotics applications is generally too vast. Transfer learning refers to the area of machine learning concerned with the deployment of algorithms trained on a source domain to perform tasks on a target domain. Hierarchical RL refers to the family of techniques where the solutions to a reinforcement learning problem, usually parametric policies, have a hierarchical structure and whose components may be deployed separately.
Presented by Juan Camilo Gamboa Higuera. Slides: pdf pptx
The purpose of this meeting is to give an overview of the different problem formulations for transfer learning in robotics applicaitons. Recommended reading material:
- Transfer Learning for Reinforcement Learning Domains: A Survey
- Reinforcement Learning with Multi-Fidelity Simulators
- Multi-Fidelity Reinforcement Learning with Gaussian Processes
- Unsupervised Cross-Domain Transfer in Policy Gradien
- Taskonomy
Presented by Melissa Mozifian. Slides: pdf
The purpose of this meeting is to dive into more promising Transfer techniques applied in robotics. Recommended reading material:
- Learning Dexterous In-Hand Manipulation
- Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
Presented by Monica Patel
I will explore more on question of "What to Transfer? How its related to Hierarchy?" Some of the What's are:
- Control Policies: Options, Skills, Universal Options, Absctract Markov Decision Processes, Policy sketches.
- Value Function: Universal Value Function Approximators.
- Reward: Little introduction on "Intrinsic Motivation" and Hierarchy in rewards.
- Task Parametrization: Hierarchy in lifelong long learning
Recommended reading material:
- Between MDPs and semi-MDPs
- CST: Constructing Skill Trees by Demonstration
- Planning with Abstract Markov Decision Processes
Presented by Auguste Lalande
I will explore some of the challenges faced at the perceptual level when doing transfer (e.g. overcoming visual differences between simulator and real world).
Recommended reading material:
Presented by Nikhil Kakodkar
This week we will be discussing the Option-Critic architecture in detail.
Recommended Reading: