Copyright 2000 by Neal R. Wagner.
Agents are computer programs that do routine tasks on one's behalf. (More inventive and active agents are also called ``bots'' now.) Agents are constructed in software, though they often control hardware devices. They are autonomous, that is, working on their own without minute-by-minute direction from their owners. As time passes they will be increasingly intelligent, meaning that they can learn from past experience and will do ever more complex tasks -- tasks that have required human intelligence. Society stands at the threshold of the era of autonomous, intelligent machines.
These agents are important for all areas of computing, from browsing the web to controlling a factory to fault detection in a network, but they are particularly significant for the surveillance technologies. Combing through databanks searching for the identity of a lawbreaker is an arduous task, best done by a computer agent. The widespread tracking and logging advocated here will require agents for any practical implementation, since human supervisors could not keep up with or coordinate all the data. Either immediately (in real-time) or after the fact, these agents will call suspicious activities to the attention of a human being. Computers will have agents monitoring each other and even themselves; the activities of these agents will be monitored as well, and so on in a circle or a regression.
People might picture an agent as a tiny robot, with clever miniature gears and wires and sensors, but the reality is much more interesting and exciting: These are software agents, electronic information. They are expensive to create initially, but replication is nearly free. The agents can be stored like any other data, on disk or CD or tape. They are easy to transmit over a network, so one can copy them into a home or office. They will run on standard hardware, for example, on personal computers. Here is the new breed of machines: logical machines created as a set of instructions to hardware, the most complicated machines ever constructed. Already they can have millions or even billions of components -- soon much more. Advanced agents are also mobile, meaning that they move between machines on a network (making copies of themselves), and they can cooperate with other agents toward a common goal. These versatile machines, with all their advantageous properties, as well as their potential for mischief, are transforming society.
Intelligent Agents
The real excitement in any discussion of agents is their possible intelligence. The computer genius Alan Turing developed an abstract concept of a machine in the 1930s before there were any real computers. In 1950, he addressed the idea of machine intelligence by imagining a remote-controlled typewriter, either connected to a human at the other end, or to a machine pretending to be a human. What came to be known as the Turing intelligence test for a machine is whether the machine could fool a human being into thinking it was human, too. Debates about the Turing Test swing back and forth, as one person argues that there is more to intelligence than typed responses, no matter how clever, and another says that the ability to write convincing answers would prove intelligence. Current agents are not remotely close to this goal yet, but they should be able to do useful work even within their limitations.
For thirty years the artificial intelligence community has predicted the imminent arrival of intelligent machines. I am not being entirely fair, and the early optimism has long since waned. Decades ago the artificial intelligence community started a debate of morality issues when dealing with intelligent machines that continues to this day. Should they be created? If created, is it immoral to terminate them? There are many similar questions, ones that will not need answers in my lifetime. But even if not truly intelligent, these complex and capable machines will be doing ever more significant tasks on behalf of their creators.
Computer scientists are now transforming the Internet from a passive network into an active entity, a collection of interacting, intelligent agents. Sun Microsystems has the current hot proposal, but whether or not Sun's specific products survive long-term, the promise of cooperating, transferable agents is here. Sun named its language for creating portable and downloadable agents ``Java.'' Now other caffeine-laden names permeate this area. In the future, to get a service from a Web site on the Internet, one will not ask questions and figure out how to obtain the service and how to use it, but will fetch an intelligent agent from the remote Web site into the local computer; the agent will proceed to deliver and carry out the service without intervention.
Or consider what a library is: a repository of information, right? One goes to a library, physically or electronically, and retrieves the desired information, perhaps with help from a librarian. I foresee a future with all the emphasis on library agents and none on the library itself. The intelligent library agent will retrieve data from anywhere: remote archives of data or databases, libraries, museums, special collections of every kind. The agent will be willing to take action: calculate answers to questions, experiment, access the real world, apply inference rules. For example, if one wants a map, the agent might fetch it from an archive, or might create it from raw data describing the region in the map, or might use an orbiting satellite to photograph the area of interest in real-time.
These agents will do anything, not just data processing and computing activities. They will run businesses and homes, governments and their agencies. They will manage personal affairs. Each home microwave oven will have its own address on the world's network and its own agent running it, protecting against theft or accidental misuse, cooperating with other agents in the home, calling for remote diagnosis of problems and for repairs, whether remote repairs or actual physical ones.
Agents for Monitoring
I propose to use extensive monitoring and surveillance of the physical world and of society's new electronic realms. Software agents should do this monitoring because there is too much of it for the humans and because the agents will do a good job, at least in their capacity and in their never-sleeping attention to detail.
An extended example concerning environmental pollution will help clarify these proposals. It is possible to catch polluters after the fact, but monitoring and detection of the earliest stages of pollution is a far better method -- to eliminate the pollution before much of it occurs.
Underground storage of gasoline provides a serious pollution problem for discussion. The U.S. uses such underground storage tanks everywhere, storing gasoline or oil or other liquids that will harm the water table in case of a leak. Gasoline floats on top of water, so a water well near a leak will bring up even more gasoline than might be expected from the size of the leak.
A pilot project carried out by Dr. Jerry Keating in San Antonio does a good job of detecting gasoline leaks at filling stations. A floating sensor in each tank accurately measures the tank's gasoline level in a manner similar to the gasoline gauge of an automobile. Other sensors access the gasoline pumps and determine the amount of gasoline removed. Using these sensors, a computer at the station can keep track of the gasoline in each tank -- the amount removed and the amount remaining, as a bank does accounting with its money. The software checks for a time when all the pumps attached to a given tank are closed down, and then measures the gasoline level in the tank. A leak shows up as a tank level lower than it ought to be.
In practice complicating factors intrude. A truck driving by will make waves in the tank and disturb the floating sensor, making an individual reading of the level inaccurate. Calculations must take the temperature into account and even the amount of gasoline vapor in the air above the liquid gasoline. Accuracy requires multiple readings and an application of a statistical analysis to the data (a regression analysis).
Using multiple readings of the level, the software can detect a leak of two-tenths of a gallon per hour in a week or so. In a month, the system can almost catch a leak rate of one-tenth of a gallon per hour. There are fancier sensors than a float that would increase accuracy and allow detection of even smaller leaks, but two-tenths of a gallon per hour is already a small rate. A larger leak, say, a gallon per hour or more, would show up immediately.
The U.S. federal government will require accurate reconciliation of gasoline amounts by 1998, but unfortunately, the method described above is just now being abandoned in favor of another that is simpler and less accurate. This other method sends pump readings and levels obtained with a dip stick off for processing and would only catch a much larger leak after several months of leaking.
A profit-oriented oil company has reasons not to want the accurate leak-checking system in place. By one estimate, perhaps a quarter of all tanks have a slow leak, and digging up tanks or otherwise repairing leaks is expensive. A leak of one-tenth of a gallon per hour is about 900 gallons per year -- a loss the company could ignore if it wished. Also, the system described here costs perhaps US $5000 to install, and there would be additional maintenance costs. One benefit to an oil company is early detection of gasoline theft. During the prototype experiments in San Antonio, a filling station manager filled his own truck's tank with sixteen gallons every Sunday afternoon when the station was closed. This theft was serious even though an unsophisticated manager stole a trivial amount, but the system would also catch important larger thefts.
The point for me is not this particular leak-detecting system, but the fact that such leak detection is feasible within a reasonable budget. This is an important social and environmental issue. Early detection of gasoline leaks is necessary to avoid later, more expensive problems with the water supply. If laws require all filling stations to detect these leaks, no station will be at a competitive disadvantage, and the loss in profits will be small, partly offset by the gain from not losing the gasoline.
Society should provide the world with monitors like those in a giant airliner, endless dials and gauges and warning lights, attended not by humans but by clever software agents, checking for all manner of problems and hazards.
Problems with Agents
I am a fan of these agents and would like to see them widely used, but I realize that their utility depends on the application area. For the crimes of stealing or leaking gasoline, agents for leak detection seem like a good solution. However, if society is not going to do anything about a problem uncovered by monitors and agents, there may be no reason to deploy them. For example, ozone and other pollutants in the air are often not effectively dealt with, and people may not need monitors to realize the extent of the problem. Even here the monitoring is useful to pinpoint the exact areas affected, while research may help determine individual contributions to the problem and the best ways to reduce it.
The previous section passed quickly over problems of accuracy, reliability, and maintenance of the monitors and sensors and of the computers running software agents. Anyone with troubles keeping a personal computer working may not consider this a trivial issue, but there are ways to spend more money for improved reliability, including the use of multiple central processing units and other redundant hardware. Also, reliability problems often come from changing software, and systems used for monitoring should not need such frequent changes.
The expense of wiring up the world with monitors is another concern, and this write-up contains no careful cost accounting. Nevertheless, the prevention measures made available by more-vigilant monitoring are bound to pay for themselves. As with other issues, society cannot afford not to monitor more extensively.
Finally, there is the danger that society may delegate to autonomous agents more than it intends. The New York Stock Exchange now has rules against automated stock-transacting agents after analysts blamed these agents for the severe 1987 market decline, when the computer-generated program trading agents issued huge buy and sell orders, faster than anyone could follow. The fear is of a future major crash -- perhaps greater than the 22.6% drop in 1987. Thus human supervision is needed for agents that might quickly initiate a cascade of actions.