Jihie Kim, Yolanda Gil
A learning system may need several methods that follow different learning strategies in order to learn how to perform complex tasks. For example, a learning method may be used to generalize from user demonstrations, another to learn by practice and exploration, and another to test hypotheses with experiments. In such an integrated learning system, there is a need for systematically coordinating the activities of the participating learning agents especially to ensure that the system creates appropriate procedural knowledge. In this paper, we describe an approach for developing a meta-level problem solver that coordinates different agents in an integrated learning system based on their capabilities and status of learning. The system is called Maven. Maven is cast on a BDI-style framework with explicit representation of learning goals, a set of plans for achieving learning goals, and high-level learning strategies that prioritize learning goals. By supporting both top-down and bottom-up control strategies, Maven supports flexible interaction among learners. The status of desired learning goals and goal achievement history enables assessment of learning progress over time. Maven is being used for coordinating learning agents to acquire complex process knowledge.
Subjects: 12. Machine Learning and Discovery; 15.3 Control
Submitted: May 5, 2008