CIF: Small: Controlled Sensing with Social Learning

摘要：这项研究解决了在受控传感和与社会传感器的信息融合方面日益增长的新技术的需求。社交传感器向社交网络学习，并通过社交网络相互作用，以估计潜在的状态--这个过程被称为社交学习。这项研究的目的是为受控传感和信息融合与社会学习开发数学模型、算法和分析。带社会学习的受控感知是在社会网络数据科学中构建产生式反馈模型和算法的第一步。类似的公式出现在多代理信号处理问题中，在这些问题中，自动决策者相互作用以实现感知目标。这项研究超越了经典的统计信号处理(从噪声测量中提取信号)，解决了多代理决策系统和信号处理算法如何相互作用的更深层次的问题。本研究的目标归入两个相互关联的主题：贝叶斯社会学习，其中考虑最快的变化检测和受控融合；以及大规模网络中的交互感知，其中考虑度分布动态、感染动力学和后验Cramer Rao界的自适应估计。这项研究涉及贝叶斯社会学习、随机控制和平均场动力学的相互作用。支持这项研究的关键统一主题是统计信号处理和受控感知。社会学习涉及个体主体的短视决策；而受控感知涉及时间范围内的随机控制。近视的社交传感器(本地决策者)和非近视的控制者(全局决策者)之间的这种相互作用导致了不寻常的行为。本研究的科学创新源于在贝叶斯估计、随机优化、弱收敛分析、格规划等方面的深入研究成果。

Abstract: This research addresses the growing need for new techniques in controlled sensing and information fusion with social sensors. Social sensors learn from and interact with each other over a social network to estimate an underlying state - the process is called social learning. The aim of this research is to develop mathematical models, algorithms and analysis for controlled sensing and information fusion with social learning. Controlled sensing with social learning is a first step towards constructing generative feedback models and algorithms in the data science of social networks. Similar formulations arise in multi-agent signal processing problems where automated decision makers interact to achieve a sensing goal. The research transcends classical statistical signal processing (which deals with extracting signals from noisy measurements) to address the deeper issue of how multi-agent decision systems and signal processing algorithms interact.The objectives of this research fall under two inter-related themes: Bayesian social learning where quickest change detection and controlled fusion are considered; and interactive sensing in large scale networks where adaptive estimation of degree distribution dynamics, infection dynamics and posterior Cramer Rao bounds are considered. The research involves the interplay of Bayesian social learning, stochastic control and mean field dynamics. The key unifying themes underpinning this research are statistical signal processing and controlled sensing. Social learning involves myopic decision making by individual agents; while controlled sensing involves stochastic control over a time horizon. This interaction between myopic social sensors (local decision makers) and a non-myopic controller (global decision maker) results in unusual behavior. The scientific innovation of this research stems from advancing deep results in Bayesian estimation, stochastic optimization, weak convergence analysis, and lattice programming.