Tractable NAT-Modeled Bayesian Networks and Privacy Sensitive Construction of Agent Organizations

(1)在个人和移动设备中，决策经常使用僵化、简单化的规则。例如，一个电话号码可能会因为报告来自该号码的垃圾电话而被放在黑名单上，从而导致来自该号码的未来呼叫被过滤。该规则忽略了报告本身可能是垃圾邮件，从而导致不良行为的可能性。贝叶斯网络(BNS)是一种基于知识的系统，能够权衡复杂的上下文信息，可以帮助用户做出更智能的决策。由于一般BN的推理是难以处理的，因此有必要识别能够实现有效推理的BN的子类。它们包括低树宽的BN和高树宽但在局部结构中编码上下文特定独立性(CSI)并被编译成算术电路或和积网络的BN。 在无阻碍噪声和树(NAT)模型中编码的因果影响独立性(ICI)与CSI是正交的，NAT模型比基于CSI的局部结构(如代数决策图)更紧凑。本研究将研究如何在高树宽的NAT模型的BN中进行易处理的推理，如何利用CSI和NAT可表达的ICI来进一步提高推理效率，以及如何通过机器学习来获得这样的BN。它的成功将拓宽高树宽的易处理BN的子类，使BN推理具有更广泛的可扩展性。 (2)协同智能系统(称为智能体)非常适合于监控复杂设备或供应链中的协同设计等应用。通常，代理通过一个组织进行合作。连接树(JT)就是这样一种组织，并且被发现优于通常使用的伪树。代理可以嵌入丰富的知识，例如关于设备子系统，其是子系统供应商专有的并且需要保持私密。然而，构建JT组织的常见方法受到了此类隐私的侵犯。其结果是，供应商面临失去知识产权的风险。 为了改善这些代理系统中的隐私，本研究研究了如何在不丢失隐私的情况下构建JT组织，以及在不可避免的情况下以最小的损失构建JT组织。还开发了具有隐私保护的灵活JT组织结构，以处理系统组成中的变化，例如，当由于系统扩展而增加代理时。将研究完全自主的、保护隐私的JT构建的可行性，例如，不使用外部指定的领导代理。这项研究的成功完成将填补在JT组织基础上建立的代理系统中隐私丢失的漏洞。强大的隐私保障，再加上JT组织其他优越的计算特性，将使这些代理系统得到更广泛的应用。

(1) Rigid, simplistic rules are often used for decision making in personal and mobile devices. For example, a phone number may be placed on a black list, due to report of a spam call from it, causing future calls from the number to be filtered. The rule ignores the possibility that the report may itself be a spam, leading to undesirable actions. Bayesian Networks (BNs), knowledge based systems capable of weighing complex context information, can aid users with more intelligent decisions. Since inference in general BNs is intractable, it is necessary to identify subclasses of BNs that enable efficient inference. They include BNs of low treewidth, and BNs of high treewidth but encoding Context-Specific Independence (CSI) in local structures and being compiled into Arithmetic Circuits or Sum-Product Networks. Independence of Causal Influence (ICI) encoded in Non-impeding noisy-AND Tree (NAT) models are orthoganal to CSI and NAT models are more compact than CSI based local structures, such as Algebraic Decision Diagrams. This research will investigate how to conduct tractable inference in NAT-modeled BNs with high treewidth, how to further improve inference efficiency by exploiting both CSI and NAT-expressible ICI, and how to acquire such BNs by machine learning. Its success will broaden subclasses of tractable BNs of high treewidth, making BN inference more widely deployable. (2) Cooperative intelligent systems (called agents) are well suited for applications such as monitoring complex equipment or collaborative design in supply chains. Often, agents cooperate through an organization. The Junction Tree (JT) is one such organization and is found superior than the often used Pseudotrees. An agent may embed rich knowledge, e.g., on an equipment subsystem, that is proprietary to the subsystem vendor and needs to remain private. However, common methods to construct JT organizations suffer from breach of such privacy. As a result, vendors risk losing intellectual properties. To improve privacy in these agent systems, this research studies how to construct JT organizations without privacy loss if possible and with the minimum loss if unavoidable. Flexible JT organization construction is also developed with privacy protection to handle changes in system composition, e.g., when an agent is added due to system expansion. Feasibility of fully autonomous, privacy protecting JT construction will be studied, e.g., without using an externally specified leader agent. Successful completion of this research will close the loop hole for privacy loss in agent systems built on JT organizations. The strong privacy guarantee, coupled with other superior computational properties of JT organizations, will make these agent systems more widely applicable.