In performing higher level cognitive tasks in a connectionist network, we have argued that it is useful to incorporate symbol processing formalisms and mechanisms directly into the network architecture. We have previously described a neural network, RuleNet [McMillan et al., 1992] that maps strings of input symbols to strings of output symbols. RuleNet discovers a functional categorization of the input symbols, as well as a compact set of symbolic rules that operate on these categories to generate the output strings. The integration of subsymbolic categorization and symbolic rule following in a single system yields a variety of important consequences. This paper explores one consequence in detail, namely a novel way in which RuleNet performs conflict resolution among rules: When a conflict arises, RuleNet exploits the varying category membership strengths of input symbols to select a rule. This dynamic, content-based conflict resolution technique contrasts with traditional, broad heuristic techniques such as selection of the most specific or least recently used rule. RuleNet also has the ability to learn a priority ranking of rules, and this ranking can interact with the category-strength selection technique in a uniform quantitative framework. This uniform framework-the combination of evidence in a connectionist system-is a powerful and elegant approach to conflict resolution.
展开▼