Learning Agents Center

Learning Agents Center
Student Walking on the Fairfax Campus

Research Vision: Agents Taught by Typical Computer Users

The Learning Agents Center conducts basic and applied research on the development of cognitive assistants that:

  • learn problem solving expertise directly from subject-matter experts,
  • support experts and non-experts in problem solving and decision making,
  • teach their problem solving expertise to students.

Major research areas include instructable agents, evidence-based reasoning, ontologies and rules, multistrategy learning with an evolving knowledge representation, graphical user interfaces, integrated logic and probabilistic reasoning in uncertain and dynamic environments, mixed-initiative reasoning, crowd problem solving, modeling experts' reasoning, learning-based knowledge engineering, natural language processing, and intelligent tutoring systems.

The Center also has the mission of supporting teaching in its areas of expertise, particularly instructable cognitive assistants, machine learning, knowledge acquisition, intelligent agents, artificial intelligence and its applications.

The Center is working toward a general theory of how subject matter experts (who do not have computer science or knowledge engineering experience) can directly develop knowledge-based agents that incorporate their expertise. The approach relies on developing powerful learning agents that can be taught by the experts in ways that are similar to how the experts would teach students or apprentices, by explaining problem solving examples, and by supervising and correcting their problem solving behavior. Because such agents learn to replicate the problem-solving behavior of their human experts, we have called them Disciple agents.

At the core of many problem solving and decision making tasks is reasoning with incomplete, ambiguous, contradictory, and inconclusive information of imperfect credibility. Therefore, the center is also working toward integrating a general computational theory of evidence-based reasoning into the Disciple learning agents. This means that, before being trained for a specific application domain, a Disciple agent already contains a significant amount of domain-independent knowledge of evidence-based reasoning, including an ontology of evidence and rules for determining the credibility of evidence. This enables more rapid development of agents for specific evidence-based reasoning applications.

In the long term, the learning and evidence-based reasoning theory and technology developed in the center (called the "Disciple approach") contributes to a new revolution in the use of computers by enabling typical computer users to develop their own cognitive assistants. Thus, non-computer scientists will no longer be only users of generic programs developed by others (such as word processors or Internet browsers), as they are today, but also agent developers themselves. They will be able to train their personal Disciple assistants to help them with their increasingly complex tasks in the knowledge society, which should have a significant beneficial impact on their work and life.

In the short term, the center follows a spiral approach toward its long term goal through successive cycles of:

  • basic and applied research on learning agents,
  • design and development of tools for building learning agents,
  • employment of these tools to develop agents for a wide variety of applications, and
  • transition of these agents to end-users.

Current and previous application domains include cybersecurity, intelligence analysis, critical thinking education, behavior modeling, military center of gravity determination, course of action critiquing, emergency response planning, financial services, medicine, as well as personalized training in these domains.