Martin Riedmiller - The Robot Learning Seminar Series скачать в хорошем качестве

Martin Riedmiller - The Robot Learning Seminar Series

The Robot Learning Seminar Series is a regular series of in-person seminars at Imperial College London, hosted by Edward Johns, Director of the Robot Learning Lab. For more information, please visit www.robot-learning.uk/seminar-series. Speaker: Martin Riedmiller (Google DeepMind) Title: Data-efficient RL Agents - how to build and why they matter Date: Wednesday 12th July 2023 Abstract: ‘Intelligence is the ability to efficiently and effectively generate knowledge out of experience’ - guided by this hypothesis, we investigate algorithms and agent architectures that can autonomously learn with minimal interactions and from minimal prior knowledge. I will discuss the ‘collect & infer’ principle that provides the blueprint for an agent architecture of interacting learning processes and give concrete examples of their implementation. The learning behaviour will be shown on several examples in the field of control and robotics. Biography: I’m a research scientist and former professor for machine learning and now a team lead at DeepMind. My core scientific interest are intelligent machines, that are able to autonomously learn new things from scratch. I’m in particular interested in neural networks - a sort of mathematical model of the brain - and their ability to store and generalize information. This fascination goes back to my master thesis about supervised learning algorithms (‘Rprop’, 1992). I have always been working at the boundary of new machine learning methods and their application to novel challenges: neural forecasting systems for financial trading and sales rate prediction (‘George’, 1994; ‘Bild-Zeitung’, 1998 - 2008), self-learning agents that control self-driving cars (at Stanford, 2006) or reading thoughts and even controlling brain activity ('BrainLinks BrainTools', 2011 - 2019). Brainstormers, our robotic soccer team, was a 5 times winner of the RoboCup World Championship and one of the first teams to use reinforcement learning (RL) as their core method (1998 - 2008). The data-efficient reinforcement learning algorithms Neural Fitted Q Iteration (NFQ, 2005; NFQCA, 2011) and Deep Fitted Q (DFQ, 2010) laid the ground for many methods in current Artificial Intelligence (AI) research. I have followed my interests in various roles: as a game programmer and author for the ZX81 and ZX Spectrum (1981-1986), a computer science professor at the universities of Dortmund, Osnabrück and Freiburg (2002-2015), a Co-Founder of one of the first startups in modern AI (Cognit - Lab for learning machines, 2010-2015). In 2015, I joined DeepMind as a research scientist and team lead of the Controls Team.