- Speaker： Dr. Chengchun Shi (史成春) @ London School of Economics and Political Science
- Title： MAIB-Talk-010: Manifold Learning and Artificial Intelligence - Statistical Inference in Reinforcement Learning
- Date：18:30pm US East time, 03/11/2023
- Date：07:30am Beijing time, 03/12/2023
- Zoom ID：933 1613 9423
- Zoom PWD：416262
- URL: https://uwmadison.zoom.us/meeting/register/tJcudu-prTIuGNda1MsF8PKyRQlnGn06TP2E
Title: Statistical Inference in Reinforcement Learning
Abstract: Reinforcement learning (RL) is concerned with how intelligence agents take actions in a given environment to maximize the cumulative reward they receive. In healthcare, applying RL algorithms could assist patients in improving their health status. In ride-sharing platforms, applying RL algorithms could increase drivers’ income and customer satisfaction. RL has been arguably one of the most vibrant research frontiers in machine learning over the last few years. Nevertheless, statistics as a field, as opposed to computer science, has only recently begun to engage with reinforcement learning both in depth and in breadth. In today’s talk, I will discuss some of my recent work on developing statistical inferential tools for reinforcement learning, with applications to mobile health and ridesharing companies. The talk will cover several different papers published in highly-ranked statistical journals (JASA & JRSSB) and top machine learning conferences (ICML)
Bio: Chengchun Shi is an Assistant Professor in data science at London School of Economics and Political Science. He is serving as the associate editors of JRSSB and Journal of Nonparametric Statistics. His research focuses on developing statistical learning methods in reinforcement learning and analysis of complex data, with applications to healthcare, ridesharing and neuroimaging. He was the recipient of the Royal Statistical Society Research Prize in 2021. He also received the IMS travel awards in three years.
https://callmespring.github.io/
报告人简介：史成春，2019年获美国北卡罗来纳州立大学博士学位，即将成为伦敦政治经济学院教授。对统计与机器学习、个性化处理、高维数据建模等有理论和应用意义的科学问题，开展了若干极有价值的研究工作。在包括《Journal of the Royal Statistical Society, Series B》、《Annals of Statistics》、《Journal of the American Statistical Association》和《Journal of Machine Learning Research,》等在内的国际顶级期刊发表论文十余篇。
Background:
Abstract: Reinforcement learning (RL) is a powerful tool for developing intelligent agents that can learn to take actions in various environments to maximize rewards. RL has the potential to revolutionize healthcare and ridesharing industries by assisting patients in improving their health and increasing drivers’ income and customer satisfaction, respectively. However, statistical inference for RL has only recently received attention from the statistics community. In this talk, I will present my recent work on developing statistical inference methods for RL with applications to healthcare and ridesharing. I will discuss several papers published in top statistical and machine learning journals and conferences, including the Journal of the American Statistical Association (JASA), the Journal of the Royal Statistical Society: Series B (JRSSB), and the International Conference on Machine Learning (ICML). The talk will cover topics such as model-based RL, off-policy evaluation, and sample-efficient RL, and their applications in mobile health and ridesharing platforms.
Bio: Chengchun Shi is an Assistant Professor of Data Science at the London School of Economics and Political Science. He is an associate editor of JRSSB and the Journal of Nonparametric Statistics. His research interests lie in developing statistical learning methods for complex data analysis, with a focus on reinforcement learning applications in healthcare, ridesharing, and neuroimaging. Dr. Shi was awarded the Royal Statistical Society Research Prize in 2021 and has received the IMS travel awards for three consecutive years.
Future Challenge in this field:
Sample Efficiency: Reinforcement learning algorithms often require a large amount of data to learn effectively. Improving sample efficiency is critical for making RL algorithms more practical and scalable to real-world applications.
Generalization: Many RL algorithms have difficulty generalizing to new environments or tasks that differ significantly from their training data. Developing algorithms that can generalize effectively is an important challenge for RL.
Safety and Robustness: As RL algorithms are increasingly applied to safety-critical domains, ensuring the safety and robustness of RL agents is of utmost importance. This requires developing methods for evaluating the safety of RL agents and designing algorithms that can guarantee safe behavior.
Explainability: As RL is used in more complex and real-world scenarios, it becomes increasingly important to understand why an RL agent takes a certain action. Developing methods for explaining the behavior of RL agents is a key challenge in the field.
Fairness: RL algorithms have the potential to exacerbate existing social inequalities if they are not designed with fairness in mind. Ensuring that RL agents behave fairly and do not perpetuate discriminatory practices is an important challenge for the field.
Multi-agent RL: Many real-world applications involve multiple agents interacting with each other. Developing RL algorithms that can effectively learn in these multi-agent settings is a significant challenge.
