The newest agents working 24 hours a day to uncover threats to homeland security aren’t human but rather pieces of extremely efficient computer code.

of the team, led by Thomas Potok at the Applied Software Engineering Research Group, is to use software agents to gather the huge amounts of information available and to cluster them according to content similarity, thereby reducing the data to something manageable that a human analyst can examine.

“The challenge is to take an incredible amount of information and very quickly determine what represents a true threat to our safety,” said Potok, who leads a team of researchers in the lab’s Computational Sciences and Engineering Division. “What we are trying to do is take 100,000 pages and bring it down to the 100 most important pages that a human can look at and make a connection. We are not making a decision for human analysts. This is just a way of reducing the information so a human can make the decision.”

Already at work

Currently, about 64 intelligent agents working on a corresponding number of computers scan the Web looking for sensitive information embedded in text at the rate of 10,000 to 15,000 pages at any given moment. Each agent is created with what Potok terms a unique piece of “DNA”—an exclusive piece of software code that gives the agent a clear directive to look for specific keywords and to bring back certain types of information. Every agent can be designed to have a slightly different DNA and be on the alert for a variety of information. Successful agents are also being designed to “mate” with each other to create the next generation of specialized agents that search more efficiently than their predecessors.

“The idea is to take the two agents that found the best information and combine that information to make a new agent,” said Potok. “Let’s say I have an agent that’s looking at terrorist groups and another that’s looking at methods that these groups may employ. If there’s a lot of activity by one terrorist group and that group is known to use a certain method, and if I find from my method agent that a certain type of anthrax is being sought after, I could then say I think there may be a link between those two. By combining the terrorist agent and the method agent, I’ll have a new agent with a more effective search parameter.”

The agents have been created to operate collectively like a community; once an agent spots sensitive information, it will leave a marker behind so that other agents can pick up the trail and examine the data more closely. Successive generations of agents will not only search for more specific information but also link them together to present the human agent with possible scenarios. The ultimate aim is to be able to detect an imminent threat or pattern that no one has been able to spot with conventional methods long before they materialize. While Potok could not comment on the results due to security reasons, he said that the agents had helped the analysts in uncovering useful information.

Dangers

“There is the very real possibility that the ‘enemy’ will attempt to understand these artificially intelligent systems and manipulate them,” said Thomas Keenan, a professor at the University of Calgary, Canada, and a fellow of its Centre for Military and Strategic Studies. “It is a classic cat-and-mouse game, but if it were possible to figure out, for example, the priorities and weighting of the system—like newspaper articles versus satellite images—a campaign of disinformation could easily be launched to weaken the power of a system like this. But there is no question that those who combat terrorism need every possible tool, and intelligent agents can at least reduce the burden of trying to keep track of large numbers of websites, some of which may come and go in a few hours.”

The ORNL team plans to work with larger computers next and explore the possibility of getting the agents to work off sensors. They face challenges that include scalability—which involves figuring out how thousands or even millions of agents can communicate with each other and people—and developing agents that more closely mimic human brain functions. The technology is still in the research phase, but the team hopes that it will be implemented in a major way in one to three years. HST