GrAM: Reasoning with Grounded Action Models by Combining Knowledge Representation and Data Mining

Abstract

This paper proposes GrAM (Grounded Action Models), a novel integration of actions and action models into the knowledge representation and inference mechanisms of agents. In GrAM action models accord to agent behavior and can be specified explicitly and implicitly. The explicit representation is an action class specific set of Markov logic rules that predict action properties. Stated implicitly an action model defines a data mining problem that, when executed, computes the model’s explicit representation. When inferred from an implicit representation the prediction rules predict typical behavior and are learned from a set of training examples, or, in other words, grounded in the respective experience of the agents. Therefore, GrAM allows for the functional and thus adaptive specification of concepts such as the class of situations in which a special action is typically executed successfully or the concept of agents that tend to execute certain kinds of actions.

GrAM represents actions and their models using an upgrading of the representation language OWL and equips the Java Theorem Prover (JTP), a hybrid reasoner for OWL, with additional mechanisms that allow for the automatic acquisition of action models and solving a variety of inference tasks for actions, action models and functional descriptions.