Chemical discovery as belief revision

摘要

In this paper we describe STAHLp, a system that constructs componential models of chemical substances. STAHLp is a descendant of Zytkow and Simon's (1986) STAHL system, and both use chemical reactions and known componential models in order to construct new chemical models. However, STAHLp employs a more unified and effective strategy for recovering from erroneous inferences, based partly on de Kleer's (1984) assumption-based method of belief revision. This involves recording the underlying source beliefs or premises which lead to each inferred reaction or model. Where Zytkow and Simon's system required multiple methods for detecting errors and recovering from them, STAHLp uses a more powerful representation and additional rules which allow a unified method for error detection and recovery. When given the same initial data, the new system constructs the same historically correct models as STAHL, but it has other capabilities as well. In particular, STAHLp can modify data it has been given if this is necessary to achieve consistent models, and then proceed to construct new models based on the revised data.