Computers are as ubiquitous as automobiles and toasters, but exploiting their capabilities still seems to require the training of a supersonic test pilot. VCR displays blinking a constant 12 noon around the world testify to this conundrum. As interactive television, palmtop diaries and "smart" credit cards proliferate, the gap between millions of untrained users and an equal number of sophisticated microprocessors will become even more sharply apparent. With people spending a growing proportion of their lives in front of computer screens--informing and entertaining one another, exchanging correspondence, working, shopping and falling in love--some accommodation must be found between limited human attention spans and increasingly complex collections of software and data.
Computers currently respond only to what interface designers call direct manipulation. Nothing happens unless a person gives commands from a keyboard, mouse or touch screen. The computer is merely a passive entity waiting to execute specific, highly detailed instructions; it provides little help for complex tasks or for carrying out actions (such as searches for information) that may take an indefinite time.
If untrained consumers are to employ future computers and networks effectively, direct manipulation will have to give way to some form of delegation. Researchers and software companies have set high hopes on socalled software agents, which "know" users' interests and can act autonomously on their behalf. Instead of exercising complete control (and taking responsibility for every move the computer makes), people will be engaged in a cooperative process in which both human and computer agents initiate communication, monitor events and perform tasks to meet a user's goals.
The average person will have many alter egos in effect, digital proxies-- operating simultaneously in different places. Some of these proxies will simply make the digital world less overwhelming by hiding technical details of tasks, guiding users through complex on-line spaces or even teaching them about certain subjects. Others will actively search for information their owners may be interested in or monitor specified topics for critical changes. Yet other agents may have the authority to perform transactions (such as on-line shopping) or to represent people in their absence. As the proliferation of paper and electronic pocket diaries has already foreshadowed, software agents will have a particularly helpful role to play as personal secretaries--extended memories that remind their bearers where they have put things, whom they have talked to, what tasks they have already accomplished and which remain to be finished.
This change in functionality will most likely go hand in hand with a change in the physical ways people interact with computers. Rather than manipulating a keyboard and mouse, people will speak to agents or gesture at things that need doing. In response, agents will appear as "living" entities on the screen, conveying their current state and behavior with animated facial expressions or body language rather than windows with text, graphs and figures.
Although the tasks we would like software agents to carry out are fairly easy to visualize, the construction of the agents themselves is somewhat more problematic. Agent programs differ from regular software mainly by what can best be described as a sense of themselves as independent entities. An ideal agent knows what its goal is and will strive to achieve it. An agent should also be robust and adaptive, capable of learning from experience and responding to unforeseen situations with a repertoire of different methods. Finally, it should be autonomous so that it can sense the current state of its environment and act independently to make progress toward its goal.
Programmers have difficulty crafting even conventional software; how will they create agents? Indeed, current commercially available agents barely justify the name. They are not very intelligent; typically, they just follow a set of rules that a user specifies. Some Email packages, for example, allow a user to create an agent that will sort incoming messages according to sender, subject or contents. An executive might write a rule that forwards copies of all messages containing the word "meeting" to an administrative assistant. The value of such a minimal agent relies entirely on the initiative and programming ability of its owner.
Artificial-intelligence researchers have long pursued a vastly more complex approach to building agents. Knowledge engineers endow programs with information about the tasks to be performed in a specific domain, and the program infers the proper response to a given situation. An artificially intelligent Email agent, for example, might know that people may have administrative assistants, that a particular user has an assistant named, say, George, that an assistant should know the boss's meeting schedule and that a message containing the word "meeting" may contain scheduling information. With this knowledge, the agent would deduce that it should forward a copy of the message.
People have been trying to build such knowledge-based agents for 40 years. Unfortunately, this approach has not yet resulted in any commercially available agents. Although knowledge engineers have been able to codify many narrow domains, they have been unable to build a base of all the common sense information that an agent might need to operate in the world at large. At present, the only effort to systematize that knowledge is at the CYC project at Cycorp in Austin, Tex. [see "Artificial Intelligence," by Douglas B. Lenat, page 62 ]. It is too early to tell whether a CYC-based agent would have all the knowledge it needs to make appropriate decisions and especially whether it would be able to acquire idiosyncratic knowledge for a particular user. Even if CYC is successful, it may prove hard for people to trust an agent instructed by someone else.
Both the limited agents now distributed commercially and the artificial-intelligence versions under development rely on programming in one form or another. A third and possibly most promising approach employs techniques developed in the relatively young field of artificial life, whose practitioners study mechanisms by which organisms organize themselves and adapt in response to their environment. Although they are still primitive, artificial-life agents are truly autonomous: in effect, they program themselves. Their software is designed to change its behavior based on experience and on interactions with other agents. At the Massachusetts Institute of Technology, we have built software agents that continuously watch a person's actions and automate any regular patterns they detect. An E-mail agent could learn by observation that the user always forwards a copy of a message containing the word "meeting" to an administrative assistant and might then offer to do so automatically.
Agents can also learn from agents that perform the same task. An E-mail agent faced with an unknown message might query its counterparts to find out, for example, that people typically read E-mail messages addressed to them personally before they read messages addressed to a mailing list. Such collaboration can make it possible for collections of agents to act in sophisticated, apparently intelligent ways even though any single agent is quite simple.
Over time, "artificial evolution" can codify and combine the behaviors of the most effective agents in a system (as rated by their owners) to breed an even fitter population. My colleagues and I have built such a system to develop agents to search a database and retrieve articles that might interest their users. Each succeeding generation matches its owner's interests better.
In time, this approach could result in a complete electronic ecosystem housed in the next century's computer networks. Agents that are of service to users or to other agents will run more often, survive and reproduce; those that are not will eventually be purged. Over time, these digital life-forms will fill different ecological niches some agents could evolve to be good indexers of databases, whereas others would use their indices to find articles of interest to a particular user. There will be examples of parasitism, symbiosis and many other phenomena familiar from the biological world. As external demands for information change, the software ecosystem will continually renew itself.
Obviously the widespread dissemination of agents will have enormous social, economic and political impact. Agents will bring about a social revolution: almost anyone will have access to the kind of support staff that today is the mark of a few privileged people. As a result, they will be able to digest large amounts of information and engage in several different activities at once. The ultimate ramifications of this change are impossible to predict.
The shape of the changes that agents bring will, of course, depend on how they are employed; many questions have yet to be answered, others even to be asked. For example, should users be held responsible for the actions of their agents? How can we ensure that an agent keeps private all the very personal information it accumulates about its owner?
Should agents automate the bad habits of their owners or try to teach them better ones (and if so, who defines "better")? As the electronic ecosystem grows in complexity and sophistication, will it be possible to ensure that there is still enough computing power and communications bandwidth left over for the myriad tasks that human beings want to get accomplished? The limited experiments that researchers have performed thus far only hint at the possibilities now opening up.
PATTIE MAES is associate professor of media arts and sciences at the Massachusetts Institute of Technology. In 1987, Maes received a doctorate in computer science and artificial intelligence from the University of Brussels. Since then, she has been working at M.I.T., first in collaboration with Rodney Brooks and Marvin Minsky on the construction of autonomous intelligent robots and, more recently, on the development of intelligent software agents that both assist and entertain. She is a founder of Agents, Inc. , in Boston,Mass., one of the first companies to commercialize software-agent technology.
INTELLIGENT AGENTS. Special issue of Communications of the ACM, Vol. 37, No. 7; July 1994.
OUT OF CONTROL: THE NEW BIOLOGY OF MACHINES, SOCIAL SYSTEMS AND THE ECONOMIC WORLD. Kevin Kelly. Addison-Wesley,1995.
M.I.T. Media Lab's World Wide Web home page: http://www.media.mit.edu/