Programming Research Group
Technical Report TR-6-99
Learning about Actions and Change: An Inductive Logic Programming Approach
Stephen Moyle
September 1998, 91pp.
Abstract
The event calculus is a formalism for common-sense reasoning about actions and change in dynamic systems. It has been used in diverse areas including planning and communications protocol specification. Writing event calculus programs requires the construction of domain specific axioms (DSAs) - a programming task which is non-trivial, and one that hinders the broader use of the event calculus. This work demonstrates that such axioms can be learned from temporal observations using Inductive Logic programming (ILP) techniques, in particular theory completion. The theory of logical back-propagation as a mechnism for theory completion is described and its implementation in the ILP system Progol is used here. These techniques were used to investigate learning DSAs for the traditional AI blocks world. In the experiments Progol, utilising logical back-propagation, learned correct DSAs. These results provide encouragement and highlight the possibility of discovering casual relationships from data in temporal databases, and also learning the domain specific knowledge necessary in the development of plans.
Keywords:
Machine Learning, Inductive Logic Programming, Common-sense reasoning, Event Calculus, Theory revision, Logical back-propagation.
This paper is available as a PostScript file.