Speaker
Description
Abstract: Reinforcement Learning has demonstrated tremendous success in many challenging tasks with superhuman performance. Nevertheless, many of the decision-making problems in network optimization/scheduling naturally involve the participation of multiple decision-making agents (e.g., a network of routers/switches and a group of decentralized controllers) and thus need to be modeled as Multi-Agent Reinforcement Learning (MARL) problems. As the number of agents grows, we start to encounter “the curse of many agents” -- the exponential growth of MARL problem space significantly hinders the development of collaborative exploration strategies as well as the learning of joint decision-making policies. In this talk, we will present our recent research on multi-agent learning algorithms and multi-agent option/skill discovery, as well as their applications to traffic engineering, scheduling, and 5G slice management.