Very Large-Scale Multi-Agent Simulation in AgentScope

最近大规模语言模型（LLMs）的进展为在大规模仿真中应用多智能体系统开辟了新的途径。然而，使用现有平台进行多智能体仿真仍然存在一些挑战，例如有限的可扩展性和低效率、智能体多样性不足以及管理过程需要花费大量精力。为了解决这些挑战，我们为AgentScope开发了几个新功能和组件，增强了其方便性和灵活性，以支持非常大规模的多智能体仿真。具体而言，我们提出了一个基于Actor的分布式机制作为底层技术基础，以实现高可扩展性和高效率，并提供灵活的环境支持来模拟各种真实场景，从而实现多智能体的并行执行、集中式工作流编排以及智能体之间和智能体与环境之间的交互。此外，我们在AgentScope中集成了一个易于使用的可配置工具和一个自动生成背景的管道，简化了创建具有多样化但详细背景设置的智能体的过程。最后，我们提供了一个基于Web的界面，方便监控和管理可能部署在多个设备上的大量智能体。我们进行了全面的仿真，以展示AgentScope中所提出的增强功能的有效性，并提供了详细的观察和讨论，以突出在大规模仿真中应用多智能体系统的巨大潜力。源代码已在GitHub上发布，以激发进一步的大规模多智能体仿真研究和开发。

AgentScope: A User-Friendly Multi-Agent Platform for Large-Scale Simulations

The paper proposes several new features and components for AgentScope to address challenges in conducting large-scale multi-agent simulations, including an actor-based distributed mechanism for scalability and efficiency, flexible environment support, an easy-to-use configurable tool for creating diverse agents, and a web-based interface for monitoring and managing agents.

The proposed enhancements in AgentScope are demonstrated through a comprehensive simulation, and the source code is released on GitHub for further research and development. The paper highlights the potential of applying multi-agent systems in large-scale simulations and the importance of addressing scalability and diversity in agent platforms.

Related work includes research on large-scale multi-agent simulations, such as the Repast and NetLogo platforms, as well as recent advances in large language models for multi-agent systems.