ABSTRACT
Introduction
Abduction is the process of generating a best explanation for a set of observations. Symbolic models of abductive reasoning tend to be far too search-intensive, whereas connectionist models have difficulty explaining higher level abductive reasoning, such as the generation and revision of explanatory hypotheses. In addition, 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.
We propose a hybrid learning model for abduction that tightly integrates a symbolic Soar model for deliberately forming and revising hypotheses with Echo, a connectionist model for implicitly evaluating explanations (Thagard, 1989). In this model, Soar's symbolic knowledge compilation mechanism, chunking, acquires rules for forming and revising hypotheses and for taking actions based on the evaluations of these hypotheses. Thus, chunking models the problem solver's shift from deliberate to automatic reasoning. To complement this, Echo learns to provide better hypothesis evaluations by acquiring explanatory strengths based on the frequencies of events from past experience. Since Echo does not have a learning mechanism, we have extended it by adding the Rescorla-Wagner (1972) learning rule.
