Ensuring reasoning consistency in hierarchical architectures
Robert E. Wray, John E. Laird · 1998
Agents often dynamically decompose a task into a hierarchy of subtasks. Hierarchical task decomposition reduces the cost of knowledge design in comparison to non-hierarchical knowledge because knowledge is simplified and can be shared across multiple tasks. However, hierarchical decomposition does have limitations. In particular, hierarchical decomposition can lead to inconsistency in agent processing, resulting potentially in irrational behavior. Further, an agent must generate the decomposition hierarchy before reacting to an external stimulus. Thus, decomposition can also reduce agent responsiveness. This thesis describes ways in which the limitations of hierarchical decomposition can be circumvented while maintaining inexpensive knowledge design and efficient agent processing. We introduce Goal-Oriented Heuristic Hierarchical Consistency (GOHHC), which eliminates across-level inconsistency. GOHHC computes logical dependencies in asserted knowledge between goals rather than individual assertions, thus avoiding the computational expense of maintaining dependencies for all assertions. Although this goal-oriented heuristic ensures consistency, it does sometime lead to unnecessary repetition in reasoning and delay in task execution. We show empirically that these drawbacks are inconsequential in execution domains. Thus, GOHHC provides an efficient guarantee of processing consistency in hierarchical architectures. Ensuring consistency also provides an architectural framework for unproblematic knowledge compilation in dynamic domains. Knowledge compilation can be used to cache hierarchical reasoning and thus avoid the delay in reaction necessitated by decomposition. An empirical investigation of compilation in hierarchical agents shows that compilation can lead to improvement in both overall performance and responsiveness, while maintaining the low design cost of hierarchical knowledge.