Hoist: A Second-Generation Expert System Based on Qualitative Physics

Hoist system for hypothetical reasoning, Wif models the functionality of all the components, thereby creating a causal model of the Mark 45 lower hoist. This model not only simulates the correct operation of the lower hoist, it also simulates the boundaries of faulty operation. The lower hoist in its actual form (a machine in the real world) is not computer accessible; a computer cannot connect to the machine, make an alteration, and observe the results. However, Hoist contains a causal model, is computer accessible, and can identify single or multiple components whose failure could explain the faulty hoist behavior. Conventional rule-based expert systems attempt to capture an expert’s opinion. A causal reasoning expert system significantly differs from this approach because it does not rely on advice. Instead, the causal expert system, which reasons from first principles, relies on a qualitative model of the physics of the device. This article explains the new approach to the postdiction process in three sections. The first section, Problems of Conventional Expert Systems, investigates several problems of conventional diagnostic expert systems and the role causal reasoning can Through the technology of expert systems, the expertise of highly skilled personnel can be automated and used to assist lesser skilled personnel in the diagnosis and repair of complex machines. Expert systems that incorporate causal reasoning represent a second-generation approach to the provision of diagnostic assistance. The technology involved performs postdiction by reasoning from first principles. This article is based on research in qualitative physics and the philosophy of causality. A new implementation vehicle for causal reasoning is described, one that embodies hypothetical or counterfactual reasoning (Roach, Eichelman, and Whitehead 1985) in a language called Wif (What IF). This language was designed specifically for modeling causeeffect relationships. The application considered here applies postdiction analysis to a model of part of a naval weapon. The maintenance aid in use, the Hoist expert system, performs diagnostic assistance for the lower hoist of a Mark 45 naval turret gun. Conventional rule-based expert systems for diagnostic advice heuristically classify the cause of failure from malfunction symptoms alone (Clancey 1985). Unlike Hoist, these systems do not provide a model of the object under Articles