Amazon Research Awards-Multiagent Reinforcement Learning

There is a critical need to develop versatile artificial intelligence (AI) agents capable of solving various complex missions. However, conventional AI systems based on centralized learning are difficult to scale up: they have limitations of the high cost of maintaining big data and large models, the inefficiency of learning each different task from scratch, and lack of reliability due to central node failures. To address these issues, we have developed various multiagent reinforcement learning frameworks, in which distributed AI agents share pertinent knowledge, learn generalized joint policies across related tasks, and coordinate with each other to achieve task objectives efficiently. Recently, we are exploring a new multiagent reinforcement learning framework based on meta-learning to enable agents to adapt fast with respect to the fellow agents’ non-stationary policies.