Title: Peking University Leads Research on Efficient Reinforcement Learning for Large-scale Multi-Agent Systems Published in Nature Machine Intelligence Sub-journal
Beijing, China – A groundbreaking research led by the Beijing University of Technology’s Artificial Intelligence Research Institute has been published in the prestigious sub-journal Nature Machine Intelligence of the renowned Nature magazine. The study, titled Efficient Reinforcement Learning for Large-scale Multi-Agent Systems, is the result of collaborative efforts by Professor Yang Yaodong’s research group and has been hailed as a significant advancement in the field of artificial intelligence.
The research paper, with the first author being doctoral student Ma Chengdong from the Beijing University of Technology’s Artificial Intelligence Research Institute and corresponding author Assistant Professor Yang Yaodong, represents a significant leap forward in the field of multi-agent systems. The study, co-first authored by researcher Li Aming from the Institute’s Multi-Agent Center and Professor Du Yali from King’s College London, introduces a novel approach to decentralized collaborative training and decision-making in large-scale multi-agent systems, greatly enhancing the scalability and applicability of AI decision models in such systems.
The research addresses the challenge of implementing efficient, scalable decision-making in large-scale multi-agent systems, which is a crucial goal in the development of artificial intelligence. Multi-agent systems primarily rely on vast amounts of interaction data between agents, leveraging significant computational resources to drive each agent to learn how to collaborate with others in executing complex tasks. The core paradigm here is multi-agent reinforcement learning.
Traditional learning paradigms, such as centralized and independent learning, have their limitations. Centralized learning requires each agent to have a global view, which increases the complexity of the algorithm and communication costs, reducing scalability in large-scale systems. Independent learning, while reducing system and algorithm complexity, often results in unstable learning processes and poor decision performance.
The research team, by decoupling the dynamic characteristics of agents at the agent level and describing the relationships between agents as a networked structure, has significantly reduced the complexity of system processing. This approach, which incorporates various topologies such as linear, circular, and mesh structures, is more general and adaptable to real-world systems.
The study also proposes a more generalized networked system model to characterize the relationship between the dynamics of the decoupled multi-agent system and the real-world system dynamics. This model provides a necessary theoretical framework and analytical tools for the research of decentralized multi-agent systems.
The team further extends the model learning theory from single-agent learning to multi-agent systems, allowing agents to independently learn local state transitions, neighborhood information value, and decentralized strategies. This approach transforms complex large-scale decision-making problems into more easily solvable optimization problems.
The research team has developed a model-based decentralized strategy optimization method that is both efficient and scalable. This method can approximate monotonically improve agent strategies even with limited local information, which is essential in environments with restricted data and information interaction.
The research has been successfully applied to scenarios such as intelligent traffic control and smart grid control, demonstrating its scalability and effectiveness. Professor Yang Yaodong, the research team leader, expressed optimism about the future, stating that the team will continue to delve into the research of multi-agent learning theory and methods, empowering emerging fields like embodied intelligence and world models, and enhancing the capabilities of intelligent systems in collaboration, prediction, and decision-making in complex dynamic environments.
The publication of this research in Nature Machine Intelligence is a testament to the groundbreaking work being conducted at the Beijing University of Technology and its potential to revolutionize the field of artificial intelligence and its applications in real-world scenarios.
Views: 0