Seminar on Computational Learning and Adaptation

  Creating Cognitive Agents Using Diagrammatic Behavior Specifications

Tolga Konik
Center for the Study of Language and Information
Stanford University

Developing cognitive agents that function autonomously and intelligently in complex environments is a difficult process that requires substantial programming expertise and development time. Our goal is to automate this process using machine learning. In this talk, I present a framework where a human expert specifies abstract scenarios describing desired behavior of an agent system using a diagrammatic storyboard-like representation. In addition to the specified scenarios, the expert uses the diagrammatic tool to communicate his/her mental reasoning about the scenarios to the learning system, for example by specifying goals or by highlighting objects that are important in a decision. The learning system interprets the behavior scenarios in the context of additional background knowledge about the task and uses an Inductive Logic Programming algorithm to generate first order rules that approximate the task-performance knowledge of the expert.

This is a mixed-initiative approach where the expert specifies the scenarios interactively with a previously learned agent program. The agent program gives immediate feedback on how it would react to the specified situations, helping the expert to generate more relevant behavior data. The learning system uses these interactively specified scenarios to further improve the agent program.
 

Return to the seminar schedule