Description
[The 74th Seminar]
Date and Time: October 11, 2023: 10:00 am – 11:00 am (JST)
Venue: Online and the Open Space at the RIKEN AIP Nihonbashi office
Language: English
Title: Successor Features Representations: Human-inspired Transfer Reinforcement Learning and its Application to Social Robotics
Speaker: Chris Reinke (Inria Grenoble)
Abstract:
A goal of AI is to design agents with the same abilities as humans, including quickly adapting to new tasks. Humans learn a variety of behaviors in an environment, for example, different paths to go to work (“path A along the main street”, or “path B through the park”, and others). Given a new task (“I need to go to the bank.”), they quickly adapt by choosing their most appropriate behavior or combining them. In Reinforcement Learning (RL), which learns such multi-step decision behaviors, the Successor Representation (SR) framework, with its recent descendants Successor Features (SF) and Successor Feature Representations (SFR), allows such adaptations. SR learns the outcomes of behaviors (policies) in terms of the resulting environment dynamics (where will my behavior lead me). Given a new task (reward function), it reevaluates how its learned behaviors would perform for it and selects the most appropriate one. Behavioral evidence indicates that humans use similar strategies, which can be interpreted as an intermediate approach between model-free and model-based processes that are usually associated with human decision-making. In this talk, I introduce the SR framework, its relation to human decision-making, and how we envision its application for Social Robotics.
Biography:
Chris is a researcher in the RobotLearn Team of Xavier Alameda-Pineda at Inria Grenoble (France), where he heads the “Learning Robot Behavior” work package of the European Horizon 2020 SPRING Project (Socially Pertinent Robots in Gerontological Healthcare). His research focus is transfer and meta-reinforcement learning methods and their application in social robotics. Before, he was a postdoc in the Inria Flowers team of Pierre-Yves Oudeyer located in Bordeaux, working on automated exploration methods of complex system behaviors. He received his PhD in 2018 about brain-inspired reinforcement learning in the Neural Computation Unit of Kenji Doya at the Okinawa Institute of Science and Technology (OIST). Before his PhD, he did a Bachelor’s and a Master’s in Cognitive Science at the University of Osnabrück (Germany).
