TalkRL: The Reinforcement Learning Podcast – Détails, épisodes et analyse

What do RL researchers complain about after hours at the bar?  In this "Hot takes" episode, we find out!  

Recorded at The Pearl in downtown Vancouver, during the RL meetup after a day of Neurips 2024.  

Special thanks to "David Beckham" for the inspiration :)  

Posters and Hallway episodes are short interviews and poster summaries.  Recorded at RLC 2024 in Amherst MA.   

Featuring:  

• 0:01 David Radke of the Chicago Blackhawks NHL on RL for professional sports  
• 0:56 Abhishek Naik from the National Research Council on Continuing RL and Average Reward  
• 2:42 Daphne Cornelisse from NYU on Autonomous Driving and Multi-Agent RL  
• 08:58 Shray Bansal from Georgia Tech on Cognitive Bias for Human AI Ad hoc Teamwork  
• 10:21 Claas Voelcker from University of Toronto on Can we hop in general?  
• 11:23 Brent Venable from The Institute for Human & Machine Cognition on Cooperative information dissemination  

  
  

Arash Ahmadian is a Researcher at Cohere and Cohere For AI focussed on Preference Training of large language models. He’s also a researcher at the Vector Institute of AI.

Featured Reference

Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs

Arash Ahmadian, Chris Cremer, Matthias Gallé, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, Sara Hooker

Additional References

• Self-Rewarding Language Models, Yuan et al 2024
• Reinforcement Learning: An Introduction, Sutton and Barto 1992
• Learning from Delayed Rewards, Chris Watkins 1989
• Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning, Williams 1992

Glen Berseth is an assistant professor at the Université de Montréal, a core academic member of the Mila - Quebec AI Institute, a Canada CIFAR AI chair, member l'Institute Courtios, and co-director of the Robotics and Embodied AI Lab (REAL). 

Featured Links 

Reinforcement Learning Conference 

Closing the Gap between TD Learning and Supervised Learning--A Generalisation Point of View
Raj Ghugare, Matthieu Geist, Glen Berseth, Benjamin Eysenbach

Ian Osband is a Research scientist at OpenAI (ex DeepMind, Stanford) working on decision making under uncertainty.  

We spoke about: 

- Information theory and RL 

- Exploration, epistemic uncertainty and joint predictions 

- Epistemic Neural Networks and scaling to LLMs 

Featured References 

Reinforcement Learning, Bit by Bit 
Xiuyuan Lu, Benjamin Van Roy, Vikranth Dwaracherla, Morteza Ibrahimi, Ian Osband, Zheng Wen 

From Predictions to Decisions: The Importance of Joint Predictive Distributions 

Zheng Wen, Ian Osband, Chao Qin, Xiuyuan Lu, Morteza Ibrahimi, Vikranth Dwaracherla, Mohammad Asghari, Benjamin Van Roy  

Epistemic Neural Networks 

Ian Osband, Zheng Wen, Seyed Mohammad Asghari, Vikranth Dwaracherla, Morteza Ibrahimi, Xiuyuan Lu, Benjamin Van Roy  

Approximate Thompson Sampling via Epistemic Neural Networks 

Ian Osband, Zheng Wen, Seyed Mohammad Asghari, Vikranth Dwaracherla, Morteza Ibrahimi, Xiuyuan Lu, Benjamin Van Roy 

Additional References  

• Thesis defence, Ian Osband 
• Homepage, Ian Osband 
• Epistemic Neural Networks at Stanford RL Forum 
• Behaviour Suite for Reinforcement Learning, Osband et al 2019 
• Efficient Exploration for LLMs, Dwaracherla et al 2024 

Sharath Chandra Raparthy on In-Context Learning for Sequential Decision Tasks, GFlowNets, and more!  

Sharath Chandra Raparthy is an AI Resident at FAIR at Meta, and did his Master's at Mila.  

Featured Reference 

Generalization to New Sequential Decision Making Tasks with In-Context Learning   
Sharath Chandra Raparthy , Eric Hambro, Robert Kirk , Mikael Henaff, , Roberta Raileanu 

Additional References  

• Sharath Chandra Raparthy Homepage  
• Human-Timescale Adaptation in an Open-Ended Task Space, Adaptive Agent Team 2023
• Data Distributional Properties Drive Emergent In-Context Learning in Transformers, Chan et al 2022  
• Decision Transformer: Reinforcement Learning via Sequence Modeling, Chen et al  2021

Pierluca D'Oro and Martin Klissarov on Motif and RLAIF, Noisy Neighborhoods and Return Landscapes, and more!  

Pierluca D'Oro is PhD student at Mila and visiting researcher at Meta.

Martin Klissarov is a PhD student at Mila and McGill and research scientist intern at Meta.  

Featured References 

Motif: Intrinsic Motivation from Artificial Intelligence Feedback 
Martin Klissarov*, Pierluca D'Oro*, Shagun Sodhani, Roberta Raileanu, Pierre-Luc Bacon, Pascal Vincent, Amy Zhang, Mikael Henaff 

Policy Optimization in a Noisy Neighborhood: On Return Landscapes in Continuous Control 
Nate Rahn*, Pierluca D'Oro*, Harley Wiltzer, Pierre-Luc Bacon, Marc G. Bellemare 

To keep doing RL research, stop calling yourself an RL researcher
Pierluca D'Oro 

Martin Riedmiller of Google DeepMind on controlling nuclear fusion plasma in a tokamak with RL, the original Deep Q-Network, Neural Fitted Q-Iteration, Collect and Infer, AGI for control systems, and tons more!  

Martin Riedmiller is a research scientist and team lead at DeepMind.   

Featured References   

Magnetic control of tokamak plasmas through deep reinforcement learning 
Jonas Degrave, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de las Casas, Craig Donner, Leslie Fritz, Cristian Galperti, Andrea Huber, James Keeling, Maria Tsimpoukelli, Jackie Kay, Antoine Merle, Jean-Marc Moret, Seb Noury, Federico Pesamosca, David Pfau, Olivier Sauter, Cristian Sommariva, Stefano Coda, Basil Duval, Ambrogio Fasoli, Pushmeet Kohli, Koray Kavukcuoglu, Demis Hassabis & Martin Riedmiller

Human-level control through deep reinforcement learning
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, Demis Hassabis 

Neural fitted Q iteration–first experiences with a data efficient neural reinforcement learning method 
Martin Riedmiller  

Max Schwarzer is a PhD student at Mila, with Aaron Courville and Marc Bellemare, interested in RL scaling, representation learning for RL, and RL for science.  Max spent the last 1.5 years at Google Brain/DeepMind, and is now at Apple Machine Learning Research.   

Featured References

Bigger, Better, Faster: Human-level Atari with human-level efficiency 
Max Schwarzer, Johan Obando-Ceron, Aaron Courville, Marc Bellemare, Rishabh Agarwal, Pablo Samuel Castro 

Sample-Efficient Reinforcement Learning by Breaking the Replay Ratio Barrier
Pierluca D'Oro, Max Schwarzer, Evgenii Nikishin, Pierre-Luc Bacon, Marc G Bellemare, Aaron Courville 

The Primacy Bias in Deep Reinforcement Learning
Evgenii Nikishin, Max Schwarzer, Pierluca D'Oro, Pierre-Luc Bacon, Aaron Courville 


Additional References   

• Rainbow: Combining Improvements in Deep Reinforcement Learning, Hessel et al 2017  
• When to use parametric models in reinforcement learning? Hasselt et al 2019 
• Data-Efficient Reinforcement Learning with Self-Predictive Representations, Schwarzer et al 2020  
• Pretraining Representations for Data-Efficient Reinforcement Learning, Schwarzer et al 2021  

Julian Togelius is an Associate Professor of Computer Science and Engineering at NYU, and Cofounder and research director at modl.ai

Featured References  
Choose Your Weapon: Survival Strategies for Depressed AI Academics

Julian Togelius, Georgios N. Yannakakis

Learning Controllable 3D Level Generators

Zehua Jiang, Sam Earle, Michael Cerny Green, Julian Togelius

PCGRL: Procedural Content Generation via Reinforcement Learning

Ahmed Khalifa, Philip Bontrager, Sam Earle, Julian Togelius

Illuminating Generalization in Deep Reinforcement Learning through Procedural Level Generation

Niels Justesen, Ruben Rodriguez Torrado, Philip Bontrager, Ahmed Khalifa, Julian Togelius, Sebastian Risi