基于云强制联合差分隐私的近似真实多智能体Optimization
Approximately Truthful Multi-Agent Optimization Using Cloud-Enforced Joint Differential Privacy
摘要:
多智能体协调问题通常需要智能体交换状态信息,以达到某些集体目标,如对最终状态值的协议。在某些情况下,机会主义代理可能为了自己的利益而欺骗性地报告错误的状态值,例如,来获取更多的共享资源出于这种情况下,本文提出了一个多代理协调框架,抑制机会主义误报的状态信息。本文主要研究多智能体协调问题,可以说是非线性规划,与不可分离的约束耦合的代理。在这种情况下,一个机会主义代理可能会倾向于扭曲问题的约束,以减少其本地成本,这正是我们试图抑制的行为。所提出的框架使用原始-对偶方法,其中代理计算原始更新,并且集中式云计算机计算对偶更新。云执行的所有计算都是以一种强制联合差分隐私的方式进行的,这会增加噪声,以稀释任何代理对其费用函数值的影响。我们表明,这种稀释阻止代理故意误报他们的状态到云,并通过误报其状态代理可以实现的可能的成本降低目前的界限。这项工作扩展了我们早期的工作,将普通差分隐私到多代理优化,我们表明,这项工作可以修改,以提供一个disincentivize误报状态的云。数值结果表明,联合差分隐私下的优化算法的收敛性。
英文摘要:
Multi-agent coordination problems often require agents to exchange state information in order to reach some collective goal, such as agreement on a final state value. In some cases, it is feasible that opportunistic agents may deceptively report false state values for their own benefit, e.g., to claim a larger portion of shared resources. Motivated by such cases, this paper presents a multi-agent coordination framework which disincentivizes opportunistic misreporting of state information. This paper focuses on multi-agent coordination problems that can be stated as nonlinear programs, with non-separable constraints coupling the agents. In this setting, an opportunistic agent may be tempted to skew the problem's constraints in its favor to reduce its local cost, and this is exactly the behavior we seek to disincentivize. The framework presented uses a primal-dual approach wherein the agents compute primal updates and a centralized cloud computer computes dual updates. All computations performed by the cloud are carried out in a way that enforces joint differential privacy, which adds noise in order to dilute any agent's influence upon the value of its cost function in the problem. We show that this dilution deters agents from intentionally misreporting their states to the cloud, and present bounds on the possible cost reduction an agent can attain through misreporting its state. This work extends our earlier work on incorporating ordinary differential privacy into multi-agent optimization, and we show that this work can be modified to provide a disincentivize for misreporting states to the cloud. Numerical results are presented to demonstrate convergence of the optimization algorithm under joint differential privacy.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。