A library of general-purpose action descriptions

An important idea in knowledge representation is that of libraries of reusable knowledge components. The goal of this research is to apply this idea to action description languages. An action language may be used to specify the effects and preconditions of actions, and serves to describe “transition systems” -- directed graphs with the vertices representing the states of an action domain and the edges representing the transitions that are caused by performing actions (or by the passage of time). Many actions can be described as special cases of other actions. (For example, pushing, carrying, going can all be described as special cases of moving things around.) However, descriptions of action domains in existing action languages describe the effects of all actions from scratch, which leads to common aspects of different domains getting reinvented over and over. In this dissertation we first developed a method for defining actions in terms of other actions, in the action language C+, a language with a rich set of features for describing action domains. This provided a theoretical basis for developing a library of general-purpose action descriptions and influenced the design of the Modular Action Description (MAD) language [Lifschitz and Ren, 2006], the semantics of which is based on C+. We extended the original MAD language in several ways, both in the syntactic dimension and in the semantic dimension, and developed an implementation of this extended language. The extended semantics not only provides new features but also addresses some shortcomings of the original semantics, which were identified during the course of our research. The implemented system was used to develop a library of basic MAD modules, each describing a group of general commonsense facts related to actions. Several action domains from the knowledge representation literature were formalized using the library of basic action descriptions. The availability of the library led to the representations being much simpler than before and also enabled us to recognize structural similarities of seemingly quite different domains.