Decentralised learning and networked communication in large populations of AI decision makers

由众多相互作用的自主决策者（“代理人”）组成的系统可能在我们的社会、经济和基础设施的许多领域发挥越来越大的作用，对灾难响应、金融市场、智能城市和能源网络、环境监测和其他网络物理系统具有潜在的好处。然而，随着人口规模的增加，计算复杂性呈爆炸式增长，这使得它们难以在现实世界中使用。“平均场博弈”（mfg）是与统计物理相关的博弈论的一个领域，它可以与机器学习相结合来解决可扩展性问题。然而，解决mfg的方法传统上依赖于在实践中不现实的理想化假设。目的和目标我希望弥合mfg的抽象理论和它们在现实世界问题中的实际应用之间的差距。特别是，我将在框架中引入代理间通信，以消除对存在单个控制器的理论技术的依赖，该控制器“操纵”所有代理。这可以在健壮性、灵活性和收敛速度方面带来好处。研究方法的新颖性ymfg仍然是一个相对未被充分探索的领域，特别是考虑到我们可能需要在现实世界中训练和部署复杂系统，例如分散学习。我将代理间通信引入MFG框架是一个新颖的贡献。与EPSRC的战略和研究领域（该项目与EPSRC的研究领域相关）保持一致，有关这些领域的进一步信息可以在http://www.epsrc.ac.uk/research/ourportfolio/researchareas/Research上找到，大型多代理系统是EPSRC几个研究领域的核心，包括“人工智能技术”、“控制工程”和“验证和正确性”。潜在应用于“信息通信技术网络和分布式系统”以及“基础设施和城市系统”等领域。是否涉及公司或合作者目前没有。

Brief description of the context of the research including potential impactSystems of numerous interacting autonomous decision makers ('agents') are likely to play an increasing role in many areas of our society, economy and infrastructure, with potential benefits for disaster response, financial markets, smart cities and energy grids, environmental monitoring and other cyber-physical systems. However, there is an explosion in computational complexity as the population size increases, making them difficult to scale for real-world usage. 'Mean-Field Games' (MFGs) are an area of game theory related to statistical physics, which can be combined with machine learning to address the scalability issue. Nevertheless, methods for solving MFGs have traditionally relied on idealised assumptions that are unrealistic in practice. Aims and ObjectivesI wish to bridge the gap between the abstract theory of MFGs and their practical usage in real-world problems. In particular, I am introducing inter-agent communication into the framework, to remove the reliance of theoretical techniques on the existence of a single controller that `puppeteers' all the agents. This can bring benefits in terms of robustness, flexibility and speed of convergence. Novelty of the research methodologyMFGs remain a relatively underexplored area, especially with regards to the desiderata we may have for complex systems to be trained and deployed in the real world, such as decentralised learning. My introduction of inter-agent communication to the MFG framework is a novel contribution. Alignment to EPSRC's strategies and research areas (which EPSRC research area the project relates to) Further information on the areas can be found on http://www.epsrc.ac.uk/research/ourportfolio/researchareas/Research into large multi-agent systems is at the heart of several EPSRC research areas, including 'AI technologies', 'control engineering' and 'verification and correctness', with potential application to the likes of 'ICT networks and distributed systems' and 'infrastructure and urban systems'. Any companies or collaborators involved None currently.