报告题目：Reinforcement Learning and Multi-agent Scenarios
Reinforcement Learning is a kind of learning strategy other than Supervised Learning and Unsupervised Learning in Machine Learning context. Because of the recent success of Reinforcement Learning on single-agent games, e.g., Atari games and Go, researchers start exploring the possibility of multi-agent scenarios. The main problems of Multi-agent Reinforcement Learning lie on the dynamic adaption of the environment (including other agents) and the communication with other agents. To evaluate the performance of dynamic adaption (saying finding a stationary state of the whole system), game theory is applied and Nash equilibrium is usually utilised as a criterion. Although the model in game theory can guarantee the convergence of the system, the strong assumption makes it useless in real applications. Communication is a method that enables cooperation or coordination by gathering the local information of agents. The main issue is the communication protocol in the real world and the selection of communication targets. For these reasons, we can see Multi-agent Reinforcement Learning is still a challenge and deserves to be studied.
Jianhong Wang is a PhD candidate studying in Imperial College London, who is interested in Reinforcement Learning and Multi-agent Learning.