面向语义Web的行动表示和推理研究

The emergence and development of the semantic Web provides excellent application scenarios for action theories, but existing action theories are not suitable for the semantic Web. The challenge lies in two aspects. On the one hand, there are lots of knowledge represented as ontologies and rules on the semantic Web, and they can't be handled together by existing action theories. On the other hand, the semantic Web is an open and uncertain environment which brings challenges to actions theories. The target of this project is to construct action theories suitable for the semantic Web, by enhancing existing action theory through four steps. Firstly, the dynamic description logic DDL(X@), which was constructed by the applicants in a previous project, will be extended to support all the description logics underlying the ontology language OWL. Secondly, some elements of logic programming will be introduced into the extended dynamic description logics, and a family of dynamic description logic programs named DDL-program will be constructed as a result. With the action theory constructed over DDL-program, both the knowledge represented as ontologies and those represented as rules on the semantic Web can be handled together. Thirdly, a mechanism named belief state or some modal operators on knowledge will be introduced into the action theory; at the same time, the description of sensing actions will be also embraced into the action theory. As a result, both the partial observability of the world and the uncertainty of the effects of actions can be handled qualitatively by the action theory. Finally, some elements of Bayesian probabilities will be introduced into DDL-program, and a family of probabilistic dynamic description logic programs named pDDL-program will be constructed as a result. With the action theory constructed over pDDL-program, both the partial observability of the world and the uncertainty of the effects of actions can be handled quantitatively.

语义Web的出现为行动理论提供了一个极好的应用场景，但目前较为成熟的行动理论并不适用于语义Web环境。针对语义Web上以本体和规则的形式呈现的大量知识，以及语义Web环境的开放性和不确定性等特征，本项目在行动理论中依次增强对本体、规则、开放性和不确定性等的支持，构建适用于语义Web环境的行动表示和推理系统。首先，对动态描述逻辑DDL(X@)进行扩展，使其兼容本体语言OWL所对应的各个描述逻辑。其次，引入逻辑程序的刻画成分，构造动态描述逻辑程序DDL-program，使其支持语义Web上由本体和规则承载的大量知识。接下来，在行动理论中引入信念状态或者关于知识的模态词，引入感知行动，实现对部分可观察性和不确定性的定性处理。最后，引入贝叶斯概率，构造概率动态描述逻辑程序pDDL-program；在pDDL-program的基础上构建行动表示和推理系统，实现对部分可观察性和不确定性的定量处理。

对行动的表示和推理是人工智能领域历史最为悠久的研究主题之一。以情景演算为代表的行动理论在实际应用中取得了很好的效果。如何进一步增强行动表示和推理系统的知识表示和推理能力是研究者面临的主要问题。语义Web的出现为行动理论提供了一个极好的应用场景，但目前较为成熟的行动理论并不适用于语义Web环境，挑战主要在于两个方面：语义Web上存在大量以本体和规则的形式呈现的知识，以及语义Web环境的开放性和不确定性等特征。本项目在行动理论中依次增强对本体、规则、开放性和不确定性等的支持，构建适用于语义Web环境的行动表示和推理系统。具体来说，本项目首先对动态描述逻辑DDL(X@)进行扩展，使其兼容本体语言OWL 所对应的各个描述逻辑；同时，针对DL-Lite、EL等轻量级的描述逻辑，对动态描述逻辑的推理机制和算法进行优化实现。定义了基于动态描述逻辑DDL(SROIQ(D)@)的规划问题，给出相应的规划求解算法，并将其用于语义Web服务组合。其次，本项目在动态描述逻辑中引入逻辑程序的刻画成分，构造出动态描述逻辑程序DDL-program，使其可以处理语义Web上由本体和规则承载的大量知识。以装配序列规划为应用背景，构建了面向装配序列规划的装配本体OWL-ASP。将描述逻辑与基于事例的推理CBR结合起来，从事例表示、相似性度量、事例修正等三个方面给出了解决方案。第三，在动态描述逻辑程序DDL-program中引入了信念状态，构建了相应的行动表示系统，实现对部分可观察性和不确定性的处理。最后，为了提高计算效率，将轻量级描述逻辑抽象为图数据，在此基础上，对图数据的压缩表示和查询进行了研究。为了提高推理能力，将描述逻辑推理机制与知识图谱表示学习机制结合起来，提出了基于动态翻译原则的推理模型。本项目在执行过程中共获得软件著作权4项；申请国家发明专利3项。在国内外重要学术刊物和重要国际会议上发表学术论文46篇，其中14篇发表在SCI源刊及中国计算机学会推荐的A类和B类国际会议上。培养硕士8名、博士1名。