Homuncular learning and rule parallelism: an application to BACON

In complex knowledge systems control knowledge comprising condition-action rules is organised into knowledge sources or system operators. The paper presents a method for organising condition-action rules into a hierarchy of autonomous agents or homunculi in accordance with the tasks to be performed. In this architecture, actions are performed by action rules, tasks by homunculi and jobs by higher level homunculi. A system in this organisation can be trained at each homuncular level by explanation-based generalisation or by nonlinear classification. The paper first introduces a formalism for organising prescriptive knowledge into a homuncular hierarchy and defines homuncular learning and rule parallelism. It then describes how this method has been applied to transform the quantitative discovery program (BACON) due to P. Langley et al. (1987), into a hierarchic homuncular system and how it is trained to perform its tasks