CIF: Small: Inference over Asymmetric Network and Data Structures

在“网络”一词可以指社交网络、电力网络、运输网络或数据网络的时代，对基于图的信息处理的兴趣已经复苏。本提案下的工作与涉及大量分布式数据的应用相关，如健康信息学、监测应用、数据网络或云计算。研究结果将使工程系统受益于一个自下而上的设计范式，涉及不太强大的代理之间的协调，以实现更高水平的认知和性能的合作代理的互连网络。这项研究将发展各种技术，使各代理人能够为共同目标集体努力，并消除其互动中可能存在的不平衡现象所产生的退化效应，例如网络中的一些代理人可能比其他代理人更了解情况;一些代理人可能有专横的倾向;一些代理人出于隐私和保密考虑可能只愿意分享部分信息;并且某些代理可能遭受损坏的数据;或者具有与组不同的目标。根据这项工作开发的结果将有利于在网络科学的战略领域培养多样化的学生。研究结果还将在网上、档案出版物和会议上广泛传播给研究界。联网应用的性能指标包括：各代理人之间的合作，以赋予故障恢复能力;代理人可能不愿意与远程中心共享数据进行处理的隐私和保密考虑;并且大量数据可能已经以分散的形式可用并且在中央位置处聚集数据的事实可能是昂贵的。这些考虑促使了强大的分布式机制，使代理合作，以达到上级推理能力的发展。许多分布式技术忽略了网络和数据结构中存在的关键不对称性。机器人群是一个值得注意的应用程序的一个例子，其中代理可以受益于调整他们的探索空间，以响应不对称条件，邻居的故障，或入侵者的可疑行为。第二个例子是使用网络学习器从大型数据库中挖掘信息，例如与健康信息学、电网、社交网络或监控应用相关的数据库。如果忽略不对称性，它们可能导致错误的推断结论和性能下降。这项研究工作利用认知能力，使网络能够对抗不对称及其有害影响。特别是，该项目寻求适应和学习机制，赋予多智能体系统实现决策过程的能力，模仿动物群体的群体反应，识别和反应的能力，霸气或侵入性行为，实现分而治之的能力，集群和劳动分工策略，以及对抗损坏数据影响的能力。

In an age when the word "network" may refer to social networks, power networks, transportation networks, or data networks, interest in information processing over graphs has met with resurgence. The efforts under this proposal are relevant to applications involving large and distributed amounts of data, as is the case with health informatics, surveillance applications, data networks, or cloud computing. The research results will enable engineering systems to benefit from a bottom-up design paradigm involving coordination among less powerful agents to achieve higher levels of cognition and performance by an interconnected network of cooperating agents. The research will develop techniques that enable agents to work collectively for a common objective and to counter the degrading effect of imbalances that may exist in their interactions such as the fact that some agents in the network may be more informed than other agents; some agents may have a domineering tendency; some agents may be willing to share only partial information due to privacy and secrecy considerations; and some agents may be subject to corrupted data; or have different objectives than the group. Results developed under this work will benefit the training of a diverse body of students in the strategic area of network science. The results will also be disseminated broadly to the research community online and in archival publications and meetings.Performance indicators in networked applications include cooperation among agents to bestow resilience to failure; privacy and secrecy considerations where agents may not be comfortable sharing data with remote centers for processing; and the fact that large amounts of data may already be available in dispersed form and aggregation of the data at a central location can be costly. These considerations have motivated the development of powerful distributed mechanisms that enable agents to cooperate to attain superior inference capabilities. Many distributed techniques ignore critical asymmetries that exist in both network and data structures. Robotic swarms are one example of a notable application where agents can benefit from adjusting their exploration space in response to asymmetry conditions, malfunctioning of neighbors, or suspicious behavior by intruders. A second example is the use of networked learners to mine information from big databases, such as those related to health informatics, power grids, social networks, or surveillance applications. If asymmetries are ignored, they can lead to erroneous inference conclusions and degraded performance. This research effort exploits cognitive abilities to enable networks to counter asymmetries and their deleterious effects. In particular, this project seeks adaptation and learning mechanisms that endow multi-agent systems with the ability to implement decision-making processes that mimic quorum responses by animal groups, the ability to identify and react to domineering or intrusive behavior, the ability to implement divide-and-conquer, clustering, and labor division strategies, and the ability to counter the effect of corrupted data.