Using an ILP Algorithm to Learn Logic Programs for Reasoning about Actions.
David Lorenzo, Ramón P. Otero · 2000
Abstract. ILP is suitable to learn action descriptions of dynamic domains. We introduce an algorithm, inspired on Progol, but adapted to cope with the particular issues of action descriptions. The algorithm is shown in an application in Cognitive Robotics. 1 Introduction We focus on learning descriptions of dynamic domains that can be characterized by logic-based formalisms of action. These formalisms aim to be a theory of dynamics, grounded on a strong mathematical foundation, where system's behaviors are viewed as appropriate logical consequences of the domain's description [8]. Recently there has been much progress in formulating theories of action, particularly in progressing from simple and/or restricted theories and iexample centered approachesj, to general and provenly correct theories that incrementally consider various specio/cation aspects such as: concurrent actions, etc. This allows to consider increasingly more complex tasks, e.g., Cognitive Robotics [10]. Some of these formalisms have a Logic Programming (LP) counterpart what makes it feasible to study the integration with Inductive Logic Programming (ILP). This work deals with a logical approach to modeling dynamical systems based on a dialect of o/rst order logic called the Situation Calculus (SC). As to learning, the apprentice is provided with a sequence of states and it must infer how properties of a domain are (directly/indirectly) aoeected by the execution of actions, or otherwise are subject to the general law of persistence.