Collaborative Research: SaTC: CORE: Medium: Foundations of Trust-Centered Multi-Agent Distributed Coordination

该项目将为在对手存在的情况下分布式协调和优化多智能体系统的以信任为中心的弹性奠定理论基础。这种弹性是通过代理通过本地通信学习邻居的可信度来实现的，这使他们能够减轻对抗性行为的有害影响。特别是，代理可以识别并隔离对手，因此代理能够维持所需的系统性能。这种弹性的自主多智能体系统很可能在未来的自动驾驶车队、自动配送系统（例如机器人和无人机）以及我们家中的物理和连接设备的部署中发挥重要作用。该方法是为有效利用网络中的随机“辅助信息”建立理论基础和分析框架，以便为多智能体优化问题的弹性提供可证明的更强的保证。恶意行为通过概率链接损坏模型来解决，该模型在攻击及其对系统的影响之间提供了重要的分离。这种分离至关重要，因为它使得能够使用统计推断技术来开发信任模型。所得模型适用于研究损坏的数据对多智能体协调和优化任务的弹性的影响。这项工作的重点是推导弹性分布式优化算法和弹性共识协议，这些协议可以容忍一半以上的网络连接是恶意的；该项目旨在放松这一经典要求。该项目的具体目标是开发分布式检测攻击、缓解攻击以及描述在对手存在的情况下可实现的性能保证的方法。该贡献是一个统一的理论，用于理解如何使用代理间通信来检测和隔离恶意代理，同时可证明地量化其对系统性能的影响。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project will develop the theoretical foundations of trust-centered resilience for distributed coordination and optimization of multi-agent systems in the presence of adversaries. The resilience is to be achieved by agents’ learning trustworthiness of their neighbors through local communications, which allows them to mitigate the detrimental impact of adversarial actions. In particular, the agents can identify and isolate the adversaries and, thus, the agents are able to sustain the desired system performance. Such resilient autonomous multi-agent systems are likely to play an important role in the future deployment of autonomous vehicle fleets, automated delivery systems (such as robots and drones), as well as physical and connected devices in our homes.The approach is to establish the theoretical foundations and analytical framework for efficient exploitation of stochastic "side information" found in the network, in order to arrive at provably stronger guarantees of resilience for multi-agent optimization problems. Malicious actions are addressed through probabilistic link-corruption models, which provides an important separation between the attack and its impact on the system. This separation is critical as it enables the development of trust models using statistical inference techniques. The resulting model is suitable for studying the impact of corrupted data on the resilience of multi-agent coordination and optimization tasks. The focus in this work is on deriving resilient distributed optimization algorithms and resilient consensus protocols that can tolerate more than half of the network connectivity being malicious; a classical requirement that this project aims to relax. Specific objectives of the project are to develop methods for distributed detection of an attack, attack mitigation, and characterization of attainable performance guarantees in the presence of adversaries. The contribution is a unified theory for understanding how inter-agent communications can be used to detect and isolate malicious agents, while provably quantifying their impact on system performance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.