Representation and Reasoning about Adaptive Interfaces

之前在自适应网站、可穿戴计算和智能用户界面方面的工作表明，这些任务对机器学习、知识表示和不确定性推理领域提出了重大挑战。该项目将利用自适应界面作为灵感和实验测试平台，解决以下核心人工智能问题。1)给定一个或多个用户的行为数据数据库，编码用户行为预测模型的最佳表示是什么？学习这种模型的最佳算法是什么？本项目将推广马尔可夫模型和动态贝叶斯网络，分别创建关系马尔可夫模型（RMMs）和动态概率关系模型（dprm）。将开发有效的推理和学习算法，并对传统的命题方法进行评估。2)呈现用户界面是一个主要挑战。本项目将扩展以任务为中心的用户界面设计的工作，从规划文献（感官行动，外生事件）的想法发展一个具有明确语义的表达任务形式主义。3)适应一个界面，它被表示为一个扩展的计划模式，需要新的方法来推理动作。除了分析因果关系结构之外，类似于部分评估的重组操作也是必要的。快速推理是这个项目的重要组成部分。一个令人满意的计划是不够的，因此本工作将使用一种结合计划长度和认知失调因素的实用新型。在方法上，该项目由六个耦合的活动组成：(1)形式化RMM和DPRM表示；(2)设计高效的粒子滤波推理方法；(3)开发基于收缩的学习算法；(4)形式化一个声明性的、基于计划的接口表示，并评估适应性示例语料库的表达性；(5)设计一套全面的适应转换和效用度量；(6)实现这些方法，将其纳入用户界面平台，并进行广泛的实验。这项研究将产生广泛的影响，因为与此同时计算机速度的巨大提高相比，用户界面的进步已经相形见绌。人工智能技术可能是最有希望利用处理能力来提高用户生产力的途径。这个项目不仅有助于改进桌面软件的用户界面，而且有助于改进可穿戴计算机的个性化信息系统的用户界面。

Previous work on adaptive websites, wearable computing and intelligent user interfaces has shown that these tasks present significant challenges to the fields of machine learning, knowledge representation, and reasoning under uncertainty. This project will address the following core artificial intelligence problems using adaptive interfaces as inspiration and an experimental testbed.1) Given a database of behavioral data for one or more users, what is the best representation for encoding a predictive model of user behavior? What are the best algorithms for learning such a model? This project will generalize Markov models and Dynamic Bayes Nets to create Relational Markov Models (RMMs) and Dynamic Probabilistic Relational Models (DPRMs) respectively. Effective inference and learning algorithms will be developed and evaluated against traditional propositional methods.2) Representing user interfaces is a major challenge. This project will extend the work on task-centered user-interface design with ideas from the planning literature (sensory actions, exogenous events) to develop an expressive task formalism with clear semantics.3) Adapting an interface, which is represented as an augmented plan schema, requires new methods for reasoning about actions. In addition to analyzing causal dependency structures, restructuring operations akin to partial evaluation will be necessary. Fast inference is an essential component of this project. A satisficing plan is not good enough, so the work will use a utility model combining plan length with a cognitive dissonance factor. Methodologically, the project is composed of six coupled activities: (1) Formalize the RMM and DPRM representations; (2) Devise efficient particle-filtering inference methods; (3) Develop learning algorithms based on shrinkage; (4) Formalize a declarative, plan-based interface representation, and evaluate expressiveness on a corpus of adaptation examples; (5) Devise a comprehensive set of adaptation transformations and a utility metric; (6) Implement the methods, incorporate in a user interface platform, and perform extensive experiments. The research will have broad impact, because progress in user interfaces has been dwarfed by the simultaneous enormous increase in the speed of computers. Artificial intelligence techniques are perhaps the most promising avenue for harnessing processing power to increase user productivity. This project will contribute to improved user interfaces not only in desktop software but also in personalized information systems for wearable computers.