A Hybrid Learning Model of Abductive Reasoning

Multicausal abductive tasks appear to have deliberate and implicit components: people generate and modify explanations using a series of recognizable steps, but these steps appear to be guided by an implicit hypothesis evaluation process. This paper proposes a hybrid symbolic-connectionist learning architecture for multicausal abduction. The architecture tightly integrates a symbolic Soar model for generating and modifying hypotheses with Echo, a connectionist model for evaluating hypotheses. The symbolic component uses knowledge compilation to quickly acquire general rules for generating and modifying hypotheses, and for making decisions based on the current best explanation. The connectionist component learns to provide better hypothesis evaluation by implicitly acquiring explanatory strengths based on the frequencies of events during problem solving.