Computer Science Department School of Information Technology and Engineering George Mason University, Fairfax, VA 22030

Disciple Toolkit

The foundation of the Disciple Toolkit is an integration of apprenticeship and multistrategy learning methods within the Plausible Version Space paradigm. This paradigm allows an expert to teach the agent in much the same way in which the expert would teach a human apprentice - by giving the agent specific examples of tasks and solutions, providing explanations of these solutions, and supervising the agent as it performs new tasks. During such interactions, the expert shares his expertise with the agent, which is continuously extending and improving its knowledge and performance abilities. These kinds of agent capabilities are achieved by a synergistic integration of several learning and knowledge acquisition methods: systematic elicitation of knowledge, empirical inductive learning from examples, learning from explanations, and learning by analogy and experimentation.

Agent Building Methodology

The methodology for building and using interactive learning agents with the Disciple Toolkit is illustrated in the figure. The main objective behind the development of this methodology was to minimize the software and knowledge engineering effort needed to build the agent.

There are three stages and three different participants in the agent's lifetime:

• First, the agent developer (software and knowledge engineer) cooperating with the domain expert customizes the Disciple Toolkit and develops domain dependent software modules. The result of this effort is the customized agent with learning capabilities. It can interact with the expert during the teaching process.
• Second, the domain expert interacts with the agent to develop its initial knowledge base and then teaches the agent how to perform domain specific tasks. This process is facilitated by various knowledge elicitation and learning capabilities of the agent. As a result, a KB agent capable to operate in a given domain is released for use.
• Third, the agent is used to perform the tasks it was taught. In some domains like, for instance, in case of the Assessment Agent, agent performance is unsupervised and the agent does no longer have any learning capabilities. In other domains like, for instance, in the case of the Design Assistant, the agent's user is a domain expert, therefore the agent performance is under constant supervision and the agent can benefit from this situation by learning from the expert's criticisms.

During agent use a need for additional teaching or redevelopment may arise. Most likely the agent's knowledge will need to be updated or the agent will need to be taught new skills. The possibility to teach the agent should minimize the need for redevelopment.

Architecture of the Toolkit

The Disciple Toolkit was designed and developed to support the agent building methodology described. The toolkit has a modular architecture that facilitates the development of customized agents for various domains. The toolkit consists of: the browsing tools, the editing tools, the learning tools, and the performance tools. All the tools rely on services provided by the Knowledge Base Manager which maintains a consistent knowledge base and is specialized to the knowledge representation used by the Plausible Version Space paradigm.

Browsing and Editing Tools

The browsing tools allow for easy and natural inspection of the current state of the agent's knowledge. In this way they support the process of eliciting knowledge from the expert and the process of apprenticeship learning from examples. The browsing tools of Disciple include: the Dictionary Browser/Editor, the Concept Browser, The Rule Browser, and the Association Browser.

The editing tools are used to create, modify, and delete the agent's knowledge elements. They should be used in synergistic cooperation with the browsing and learning tools to teach the agent according to the Plausible Version Space paradigm. The editing tools of Disciple include: the Dictionary Browser/Editor, the Concept Editor, and the Rule Editor.

Learning Tools

Descriptions of the knowledge elements can be elicited from the expert with the support of the browsing and editing tools. The process of introducing directly rule descriptions to the agent's knowledge seems to be easy from the agent's point of view - it is just told what to do in certain situations. It is however not so easy from the expert's point of view. To help the expert to better communicate his knowledge about tasks and their performance the toolkit is equipped with the tools which allow learning general rules of agent's behavior from specific examples. The learning tools include: the Rule Formulation Tool, the Rule Refinement Tool, and the supporting tools like: the Example Editor, the Explanation Grapher, and the Exception Handler.

Last updated on 07/13/01