Experiments in Non-Monotonic Learning

Michael Bain

Abstract: Logic continues to have a significant role throughout AI. However, it has long been proposed that for real-world problems classical logic is unsatisfactory (e.g., where nonmonotonic reasoning may be required). The construction of incremental learning systems is a case in point. A technique called Closed-World Specialisation was recently developed to address the problem of correcting first-order theories within a non-monotonic framework for incremental learning. In this paper we report on experiments combining this technique with two methods of generalisation in first-order logic. A new inductively generated solution giving 100% predictive accuracy is presented for the task of learning rules of illegality for the KRK chess end-game.