Skill Refinement through Competence Feedback

Learning is generally performed in two stages, knowledge acquisition and skill refinement. Developments within machine learning have tended to concentrate on knowledge acquisition as opposed to skill refinement. In this paper we develop mechanisms for skill refinement within the context of lazy learning by incorporating competence feedback into the exemplar base. The extent of competence feedback depends on the richness of the data collected during task execution. We present two techniques for competence feedback, exception spaces and knowledge intensive exception spaces (KINS). The techniques differ in the extent of competence feedback and the resulting degree of skill refinement. Previous definitions of exception spaces and KINS are extended and the resulting improvement is evaluated using six data sets. A genetic algorithm is utilised to optimise the definitions of the exception spaces and KINS for each exemplar in the exemplar base. We also provide a visualisation of KINS and exception spaces with an aim to exemplify how these mechanisms for skill refinement affect the structure of the exemplar base.