Building Agents That Learn by Observing Other Agents Performing a Task: A Sequential Pattern Mining Approach

摘要

In this paper, we propose to build agents that learn by observing other agents performing a task by extracting frequent temporal patterns from their behavior. We propose a learning mechanism consisting of three phases: (1) recording other agents' behavior, (2) mining temporal patterns from this data and (3) utilizing the resulting knowledge. We illustrate this approach with a tutoring system for training learners to robotized arm manipulation where we have integrated a tutoring agent that observes humans performing a task to learn it. The agent then exploits this knowledge to provide assistance to learners.