Talk 1: “A Perspective on Safety, Risk, and Reproducibility in Reinforcement Learning”
ABSTRACT
Reinforcement learning is a field that has gained salience in recent years, dating back to the 1990s with backgammon-playing bots, and more recently with AlphaGo and training mechanisms for large language models. Reinforcement learning may refer to a problem, a learning paradigm, a collection of techniques, and so on—but irrespective of its designation, it is rife with disparate sources of randomness. We provide some perspective on how communities—spanning operations research, computer science, electrical engineering, and statistics—have tried to quantify and mitigate them. These mitigation strategies include safety, learning with constraints, risk measures, reproducibility, simulator calibration, and data coverage concerns. Time permitting, we’ll cover some emerging techniques to enforce reproducibility constraints into the standard RL training pipeline.
SPEAKER BIOGRAPHY
Alec Koppel is a senior professional staff member at the Johns Hopkins University Applied Physics Labs within the Artificial Intelligence/Machine Learning Group in the Research and Exploratory Development Mission Area. Previously, he was an AI research lead in the Multiagent Learning and Simulation Group at JPMorganChase Artificial Intelligence Research. He was also a research scientist in optimal sourcing systems within Amazon’s Supply Chain Optimization Technologies team and a research scientist at the U.S. Army Research Laboratory Computational and Information Sciences Directorate from 2017 to 2021. Koppel’s research focuses on optimization and machine learning, spanning applications in autonomous systems/robotics, financial networks, and supply chain optimizations, with particular interests in reinforcement learning/bandits, scalable online Bayesian and nonparametric methods, and online learning and stochastic optimization.
Talk 2: “Reliable Reinforcement Learning for Robot Autonomy”
ABSTRACT
Pratap Tokekar will review recent trends in safe and reliable reinforcement learning for the control of robotic autonomous systems. The Robotics Algorithms & Autonomous Systems Lab designs algorithms and builds systems to enable teams of robots to act as sensing agents, with research at the intersection of theory and systems motivated by real-world applications to environmental monitoring, infrastructure inspection, and precision agriculture.
SPEAKER BIOGRAPHY
Pratap Tokekar is an associate professor in the Department of Computer Science at the University of Maryland and an Amazon Scholar at Amazon Robotics. Between 2015 and 2019, he was an assistant professor in the Bradley Department of Electrical and Computer Engineering at the Virginia Polytechnic Institute and State University. Previously, Tokekar was a postdoctoral researcher at the University of Pennsylvania General Robotics, Automation, Sensing, & Perception Lab. He obtained his PhD in computer science from the University of Minnesota in 2014 and his bachelor of technology degree in electronics and telecommunication from the College of Engineering, Pune in India in 2008. Tokekar has received a 2022 Amazon Research Award, a 2020 NSF CAREER Award, and a 2016 NSF Computer and Information Science and Engineering Research Initiation Initiative award. He has also served as an associate editor for IEEE Transactions on Robotics and on the editorial boards for the IEEE International Conference on Robotics & Automation and the IEEE/RSJ International Conference on Intelligent Robots and Systems.