Seminar on Computational Learning and Adaptation

  Reinforcement Learning in Multiagent Systems

Junling Hu

Talkai, Inc.

In decentralized multiagent systems where agents pursue independent goals, learning plays an important role for improving each agent's performance. However, the task of learning in such environments is much more formidable than that in single-agent systems. This is mainly because other agents are learning at the same time, which causes the constant change of the learning landscape. A successful learning method therefore has to take such joint learning into account. In this talk I will focus on one learning method we propose, which allows an agent to learn its optimal actions based on its observations of joint actions and its projections of other agents' future learning. We have applied this method to several domains including E-commerce, where two online companies engage in price competition. I show that our multiagent reinforcement learning method gives considerably better performance than using single-agent reinforcement learning.

Return to the seminar schedule