Donald Rose, Pat Langley
In this paper we describe the STAHLp system for inferring components of chemical substances - i.e., constructing componential models. STAHLp is a descendant of the STAHLp system (Zytkow and Simon, 1986); both use chemical reactions and any known models in order to construct new models. However, STAHLp employs a more unified and effective strategy for preventing, detecting, and recovering from erroneous inferences. This strategy is based partly upon the assumption-based method (de Kleer, 1984) of recording the source beliefs, or premises, which lead to each inferred belief (i.e., reaction or model). STAHL’s multiple methods for detecting and recovering from erroneous inferences have been reduced to one method in STAHLp, which can hypothesize faulty premises, revise them, and proceed to construct new models. The hypotheses made during belief revision can be viewed as interpretations from competing theories; how they are chosen thus determines how theories evolve after repeated revisions. We analyze this issue with an example involving the shift from phlogiston to oxygen theory.