Agent Learning in the Emergence of Complex World
- Yali Du, University College London
- Time: 2020-04-05 15:40
- Host: Prof. Yizhou Wang
- Venue: Online Talk
Over the past few years, we have witnessed great success in AI across many applications, including image classification, recommendation systems, etc. This success has followed a common paradigm in principle: learning from static datasets with inputs and outputs. Nowadays, we are experiencing a paradigm shift from pattern recognition to decision making. Instead of learning knowledge from static datasets, we are learning through feedback on our knowledge. Especially since machine learning models are deployed in the real-world, these systems start to impact each other, turning their decision-making into a multi-agent problem. Therefore, it is essential to develop AI systems that can effectively and reliably collaborate in various contexts. This is a fundamental problem for the next generation of AI, aiming to empower various multi-agent environments.
As case studies, we present GridNet, which can flexibly control an arbitrary number of agents, and LIIR, which generates diversified behaviors when receiving only a team reward from a cooperative multi-agent system. Our novel methods achieve new state-of-the-art results on the testbed of StarCraft II environments, which has recently emerged as a challenging RL benchmark task with high stochasticity, a large state-action space, and delayed rewards. In the end, I will discuss some future directions in multi-agent learning.
Dr. Yali Du is currently a postdoctoral research fellow at University College London. She obtained her Ph.D. in AI in 2019 from University of Technology Sydney. Her research interest lies in how to address multi-agent collaboration problems, including controlling with flexibility and diversity, the emergence of interaction, multi-agent credit assignment, and robustness of agent learning. Her research output has been widely published in prestigious venues including ICML, NeurIPS, IJCAI, ACM MM, IEEE TMM, etc.