Disciple-COA: From Agent Programming to Agent Teaching

This paper presents Disciple-COA, the most recent learning agent shell developed in the Disciple framework that aims at changing the way an intelligent agent is built: from “being programmed ” by a knowledge engineer, to “being taught ” by a domain expert. Disciple-COA can collaborate with the expert to develop its knowledge base consisting of a frame-based ontology that defines the terms from the application domain, and a set of plausible version space rules expressed with these terms. Its central component is a plausible reasoner that can distinguish between four types of problem solving situations: routine, innovative, inventive and creative. This ability guides the interactions with the expert during which the agent learns general rules from specific examples, by integrating a wide range of knowledge acquisition and machine learning strategies, including apprenticeship learning, empirical inductive learning from examples and explanations, and analogical learning. Disciple-COA was developed in the DARPA's High Performance Knowledge Bases program to solve the challenge problem of critiquing military courses of action that were developed as hasty candidate plans for ground combat operations. We present the course of action challenge problem, the process of teaching Disciple-COA to solve it, and the results of DARPA’s evaluation in which Disciple-COA demonstrated the best knowledge acquisition rate and problem solving performance. We also present a separate knowledge acquisition experiment conducted at the Battle Command Battle Lab where experts with no prior knowledge engineering experience succeeded to rapidly teach Disciple-COA to correctly critique courses of action