12.2 Agents

This section introduces the agent concept and its application to navigational tasks. It is impossible to give in-depth coverage to this field in a few pages, as this is an ongoing area of research. Rather, we try to briefly illustrate the state of the art and to outline the potential of agent technology for the exploration of cyberspace.

The use of agents as a new interface model for human-computer interaction was first proposed by Alan Kay. He also called agents "soft robots".

"The model building capabilities of the computer should enable mindlike processes to be built and should allow designers to create flexible "agents". These agents will take on their owner's goals, confer about strategies (asking questions of users as well as answering their queries) and, by reasoning, fabricate goals of their own."

Allan Kay, "Computers, Networks and Education"
Scientific American, Sept. 1991, p. 148.[Kay91]

In the literature three slightly redundant different types of agents are distinguished:

• Autonomous agents are programs that travel between sites, deciding themselves when to move and what to do (e.g. General Magic's Telescript or Agent-Tcl agents). They can only travel between servers that are set up to handle agents, and are currently not widespread in the Internet.
  
• Intelligent agents are programs that help users interactively to accomplish a task, such as choosing a product, or guiding a user through form filling, or do a search. The wizards in Windows 95 are simple versions of such agents. Their intelligence is in fact quite limited, as they are best at doing a well-defined task in a limited domain. As with most AI applications, vaguely defined tasks requiring common sense knowledge are ill-suited for intelligent agents.
  
• http://www.qualcomm.com/ User-agents are programs that perform mostly networking tasks for a user, such as e-mail user-agents like Qualcomm's Eudora that filter a user's mail based on her/his interests.

Table I.3 contains a partial listing of possible tasks that can be done by an agent on behalf of the human user.

Table I.3 Types of tasks suitable for an agent (from [Lau90])

One obvious use of the agent concept is for navigation and localization of information. Contrary to guides, agents are autonomous software entities that make choices and execute actions on behalf of the user. They incorporate the knowledge to find and present information by responding dynamically to the user's changing goals, preferences, learning style and knowledge. To succeed in this task, agents need some sort of intelligence. In particular, they require knowledge about the structure and contents of the underlying information. This leads to the conclusion that structured information as, e.g., in AI knowledge bases like CYC, is particularly well suited for being managed by agents. For agents to grant access to unstructured information, this information has to be structured at least internally. For the agent prototypes available so far, most of the document structure has either been generated by manually structuring unstructured documents, or by limiting the functionality of agents to semistructured information (see the next section).

The behavior and properties of an agent must be made visible to the user. For some agents, a textual representation is completely sufficient - in Weizenbaum's ELIZA program [Wei76] its bodiless phrases may have been its greatest strength. Frequently, agents that have human-like traits are particularly well suited to make the internal properties of the agent obvious to the user. Dramatic characters are better suited than full-blown personalities, because they selectively only represent those traits that are important for the tasks the agent is supposed to support [Lau90]. Figure I.65 shows two representations of Phil, a semi-intelligent agent who appeared in the famous Knowledge Navigator videotape of Apple Computer. Although the two representations differ, the characteristic bow tie makes Phil recognizable in both representations.

Figure I.65 Phil, a fictive agent for Apple's Knowledge Navigator videotape (ExFig. 2 from [Lau90])

Compared with agents, guides are more restricted and less autonomous (table I.4). The guide idea stresses the human-like representation, but contrary to agents, guides do not contain real content knowledge about the information they manage.

Table I.4 Agents versus guides

An agent is more powerful than a guide, but unfortunately the software technology to implement general purpose agents is not yet commonly deployed. Although the foundations for implementing agents are in place, there are still many technical problems that need to be solved. In the meantime we can hardwire guides to show agent-like behavior. The end-user will not be able to notice the difference as long as the underlying knowledge base remains static and is not modified by the user.

Figure I.66 Human interacting with a virtual creature in MIT Media Lab's ALIVE project http://alive.www.media.mit.edu/projects/alive

The ALIVE project at the MIT Media Lab (figure I.66) offers a new level of immersion to the user employing virtual reality (VR) technologies. The user can interact with autonomous, intelligent agents in the form of virtual creatures (such as the dog in figure I.66) without being constrained by VR devices such as headsets, goggles, or special sensing equipment. The system is based on the so-called "magic mirror metaphor": ALIVE users see their own image on a large TV screen as if in a mirror. Autonomous, animated characters join the user's own image in the reflected world.

The ALIVE project is one of the most advanced agent projects with respect to the human computer interface. The next section introduces agents based on a textual interface, that are used to manage semistructured information. The Oval agents provide navigational cues based on semistructured content information.