Multiagent uncertain reasoning and knowledge discovery with graphical models

由于当今软件的规模和复杂性以及对可靠性的要求，以及计算机网络的进步和处理器成本的降低，通过一组称为agent的协作计算机来解决大范围内的复杂问题已经成为必要和可能。代理的一个常见任务是确定其领域的状态，以便它们采取相应的行动。我将研究执行这种推理的不同方法之间的根本区别：紧密耦合，其中代理交换有限的信念，松散耦合，其中代理将来自另一个代理的所有内容视为观察。构建这样的代理对软件开发人员来说是一个挑战。我将研究自动化一些构建步骤的方法，例如，自动增强代理接口以提高推理效率。这类代理的一个应用领域是供应链中的协同设计。没有一种现有的方法能使智能体有效地构建最优设计。继续上一节课的学习，我将分析agent有效计算最优设计的条件，并开发出一种方法。我所研究的智能体具有用概率图形模型表示的知识。从数据中学习这样的模型是自动化代理开发的一种方法。我的研究发现，有一类模型，称为PI模型，不能用常用的方法学习。如果智能体配备了一个不正确的学习模型，它可能会在执行任务时出错。为了有效地学习这些模型，需要深入了解它们是如何由独立参数组成的。对于PI模型的一些子类，我在上一阶段的研究中已经有了这样的认识。我将进一步研究，以实现对整个PI模型类的透彻理解。一旦了解了PI模型的整个频谱的组成，就可以开发新一代的学习方法，这将提高从学习模型构建的智能体的可靠性。

Due to the scale and complexity of today's software and the reliability required, and due to the advance in computer network and reduction in processor cost, it has become both necessary and possible to solve a complex problem in a large domain by a set of cooperating computers, called agents.  A common task of agents is to determine what is the state of their domain in order for them to act accordingly. I will study fundamental difference between alternative ways to perform such reasoning: tightly coupled, where agents exchange limited beliefs, and loosely coupled, where an agent treats everything from another agent as observations.  Constructing such agents is challenging to software developers. I will investigate ways to automate some steps of construction, e.g., to enhance agent interfaces automatically for improved reasoning efficiency. One application area of such agents is collaborative design in supply chains.  No existing method enables agents to construct an optimal design efficiently.  Continuing my progress in the last period, I will analyze the conditions under which an optimal design can be efficiently computed by agents and develop such a method. The agent that I study is equipped with knowledge represented as a probabilistic graphical model.  Learning such a model from data is one way to automate agent development.  My research found that a class of models, called PI models, cannot be learned by commonly used methods.  If an agent is equipped with an incorrectly learned model, it may make mistakes in carrying out its tasks.  In order to learn such models effectively, a deep understanding is needed on how they are composed from independent parameters. Such understanding on some subclasses of PI models has been obtained in my last period of research. I will further investigate to achieve thorough understanding for the entire class of PI models.  Once the composition of the entire spectrum of PI models is understood, a new generation of learning methods can be developed which will improve the reliability of the agents built from the learned models.