Model-Based Learning of Rules for Error Diagnosis

Abstract

For robot applications, automatic error diagnosis becomes an ever more important capability. Diagnosis from first principles (DFP), a technique previously applied mostly to fault diagnosis of digital circuits, promises to be a fruitful approach. Unfortunately, this technique normally fails to meet the real time requirements of robot applications. It is shown how this problem can be overcome by using explanation-based generalization (EBG). The proposed technique uses the diagnosis and its explanation from a DFP-based system and generates an error diagnosis rule. After an initial learning phase, the DFP-based system can be replaced by a rule-based diagnosis system utilizing the rules learned by EBG. This replacement can speed up the diagnosis process considerably without loosing relevant diagnosis power. The proposed combination of DFP and EBG is not limited to robot applications. It is applicable whenever DFP can be used in a given domain.