Next: A Single User's Up: Collaborative Interface Agents Previous: Collaborative Interface Agents

Introduction

Learning interface agents are computer programs that employ machine learning techniques in order to provide assistance to a user dealing with a particular computer application. Although they are successful in being able to learn their user's behavior and assist them, a major drawback of these systems is the fact that they require a sufficient amount of time before they can be of any use. A related problem is the fact that their competence is necessarily restricted to situations similar to those they have encountered in the past. We present a collaborative framework to help alleviate these problems. When faced with an unfamiliar situation, an agent consults its peers who may have the necessary experience to help it.

Previous interface agents have employed either end-user programming and/or knowledge engineering for knowledge acquisition. For example, [Lai, Malone, & Yu 1988] have ``semi-autonomous agents'' that consist of a collection of user-programmed rules for processing information related to a particular task. The problems with this approach are that the user needs to recognize the opportunity for employing an agent, take the initiative in programming the rules, endow this agent with explicit knowledge (specified in an abstract language), and maintain the rules over time (as habits change etc). The knowledge engineered approach on the other hand, requires a knowledge engineer to outfit an interface with large amounts of knowledge about the application and the domain and how it may contribute to the user's goals. Such systems require a large amount of work from the knowledge engineer. Furthermore, the knowledge of the agent is fixed and cannot be customized to the habits of individual users. In highly personalized domains such as electronic mail and news, the knowledge engineer cannot possibly anticipate how to best aid each user in each of their goals.

To address the problems of the rule-based and knowledge-engineered approaches, machine learning techniques have been employed by [Dent et al.1992][Hermens &Schlimmer1993][Maes &Kozierok1993][Kozierok &Maes1993] and others. In the Calendar Agent [Kozierok &Maes1993], memory-based reasoning is combined with rules to model each user's meeting scheduling habits. Results described in [Kozierok &Maes1993] show that the learning approach achieves a level of personalization impossible with knowledge engineering, and without the user intervention required by rule-based systems. It is also interesting to note that the addition of rules provides the flexibility to explicitly teach the agent, and shows that the rule-based and learning approaches can successfully coexist.

While the learning approach enjoys several advantages over the others, it has its own set of deficiencies. Most learning agents have a slow `learning curve' ; that is, they require a sufficient number of examples before they can make accurate predictions. During this period, the user must operate without the assistance of the interface agent. Even after learning general user behavior, when completely new situations arise the agent may have trouble dealing with them. The agents of different users thus have to go through similar experiences before they can achieve a minimal level of competence, although there may exist other agents that already possess the necessary experience and confidence.

We propose a collaborative solution to these problems. Experienced agents can help a new agent come up to speed quickly as well as help agents in unfamiliar situations. The framework for collaboration presented here allows agents of different users, possibly employing different strategies (rule-based, MBR, CBR, etc.) to cooperate to best aid their individual users. Agents thus have access to a much larger body of knowledge than that possessed by any individual agent. Over time agents learn to trust the suggestions of some of their peers more than others for various classes of situations. Thus each agent also learns which of its peers is a reliable `expert' vis-a-vis its user for different types of situations.



Next: A Single User's Up: Collaborative Interface Agents Previous: Collaborative Interface Agents


MIT Media Lab - Autonomous Agents Group - agentmaster@media.mit.edu