Ensuring Reasoning Consistency in Hierarchical Architectures

Abstract

Ensuring reasoning consistency in agents employing hierarchical task decompositions can be difficult. We have identified two specific problems that result when hierarchical context changes during problem solving. First, the agent can be too responsive, taking an external act or making an internal derivation before higher context is fully elaborated. Second, the agent can fail to respond to a change in the context, leaving a local level "unsituated " with respect to the higher contexts. In both cases, an agent’s assertions of knowledge in a local level can be inconsistent with the larger context, leading potentially to irrational behavior. Previous methods for maintaining "cross-level " consistency have relied upon a combination of both architectural techniques and explicit knowledge. Architectural