Version 0. Major revision will be available in February 2002
"Our research group seeks to help people take advantage of information. With the advent of the Internet, the World Wide Web, portable computers, portable communication devices, and greater computing power in general, the amount of information that a person confronts each day has grown astronomically. While one would hope that this information would assist people in their day-to-day lives, its sheer volume often serves to confuse and paralyze instead. Our research group develops ways to help people understand information via user interface design, information visualization, and software agency." - Information Interface Group
There is a growing sentiment that the use of agent techniques in computer interfaces is not only desirable, but necessary, due to growing volumes of information being handled. There is also strong negative sentiment toward such interfaces, however, particularly because of issues such as control and trust.
We believe that the true challenge lies in designing the right agent user interface, particularly the appropriate way to present the agent. The user-agent collaboration can only be fruitful if the user can understand what the agent does, how it does and why. And only if users know what the agent's limitations are, then they can trust the agent. Although the user turns over some control of tasks to the agent, the user must feel in control or have control whenever s/he wants it. The agent serves not as an indispensable intermediary between the user and the application, but as an adjunct to conventional interface and function as assistant rather than replacement.
One controversial aspect of agent based UI is the issue of anthropomorphism. Pattie Maes, a famous agent expert, defined anthropomorphism as "the use of graphical and/or audio interface components that give the agent a human-like representation".
We believe that rendering an computer agent human-like will make it more engaging, entertaining, approachable, and understandable to the user. It would harbor potential to build trust and establish relationships with users, and make the users feel more comfortable with computers. In face-to-face interaction, people communicate not only through words, but also by intonation, gaze patterns, facial expressions, and hand gestures. All these nonverbal behaviors are able to convey additional conversational and emotional signals and are highly linked with one another.
However, some argue that anthropomorphic agents strengthen the user’s impression of the agent’s responsibility and rationality, and thus may mislead users into thinking that the computer system is more capable than it truly is. So even if personified agents are useful in some application domain, a simple cartoon figure is enough and using an obviously non-realistic depiction for the agent is a better choice than a realistic talking head. But we believe cartoon depiction is not the only choice and in many cases not the better one. It is relatively easier for users to communicate with a unfamiliar human face rather than with never-seen cartoon character. Who would know what to say to a live paper clip when first encountered, what it could do, will behave and respond? Although a cartoon character may be more fun to play with, the goal of a software agent is to help people with their task, have work done. Also a cartoon character like the Microsoft Paper Clip may easily end up with an interface that is silly or unintentionally annoying. An agent should make every effort to build user's trust, but who will trust a cat or dog to tell them how to invest money on the market. Our goal is not, however, to build a a computer agent that looks like human, but to build a computer agent that users are able to comfortably communicate with just like human beings.
We are investigating anthropomorphic user interface metaphors through simulation, i.e. "Wizard of Oz" experiments. We seek to avoid user impressions being dictated by the competence (or lack thereof) of the system. Rather, we want to examine issues of appearance, behavior and task to learn if such interfaces have the potential to be useful and valuable.
Since new knowledge and insight are often achieved by first focusing on concrete cases and then generalizing, we direct our research to start with empirical study. From this beginning, we try to generalize to some heuristic rules for agent designers to apply.
In one sentence: to overcome the obstacles presented by speech recognition, natural language understanding, and domain knowledge.
Bibliography