Extending Explanation-Based Learning: Failure-Driven Schema Refinement

摘要

Current explanation-based learning systems assume domain theories that are computationally tractable. This paper describes a system being developed that refines schemata for use in narrative understanding a domain in which a complete analysis of agent interactions is computationally intractable. This system employs an incremental approach that learns an initial schema using the assumption that other agents will not counter-plan (i.e. take actions that will interfere with the original planners actions). However, when the system observes the failure of an actor's schema due to counter-planning by another agent, it refines the original schema. This is accomplished by indexing the counter-plan under the connecting causal chain to the original schema. This new knowledge allows the system to explain both similar failures and actions taken to prevent similar failures. This paper describes the need for incremental explanation-based learning and outlines an application of this approach to learning schemata for natural language processing.