Using Knowledge of Cognitive Behavior to Learn from Failure

When learning from reasoning failures, knowledge of how a system behaves is a powerful lever for deciding what went wrong with the system and in deciding what the system needs to learn. A number of benefits arise when systems possess knowledge of their own operation and of their own knowledge. Abstract knowledge about cognition can be used to select diagnosis and repair strategies from among alternatives. Specific kinds of self-knowledge can be used to distinguish between failure hypothesis candidates. Making self-knowledge explicit can also facilitate the use of such knowledge across domains and can provide a principled way to incorporate new learning strategies. To illustrate the advantages of self-knowledge for learning, we provide implemented examples from two different systems: A plan execution system called RAPTER and a story understanding system called Meta-AQUA.