FEED Magazine
Gift for the Gab
Clay Shirky on new software agents that evolve language

"Intelligent agents," that venerable CompSci concept, has long been something of an oxymoron. We've had a decade's worth of promises that an army of self-contained electronic processes would soon be scouring the network on our behalf, ferreting out information and finding the lowest price for everything from CDs to airline tickets. The reality, of course, is that agents have never yet found any practical application. Dr. Lee Giles and Dr. Kam-Chuen Jim, researchers at NEC, may have found the secret to transforming experimental agents into practical tools: talking helps.
What Giles and Jim have done is to solve a particularly hard problem known as Predator-Prey Pursuit, where four agents have to chase their virtual prey through a checkerboard world without knowing one another's positions. Because the agents can only capture their prey by surrounding it on four sides, group communication is the obvious tool. As Giles and Jim put it, "If an agent's [internal] state helps determine its behavior, communication may be instrumental in allowing the agents to converge on an optimal plan of action," an observation familiar to anyone who has tried to get a group to pick a movie. Until now, group communication was itself an unsolved problem. As anyone who has asked Jeeves knows, computerized grasp of human language is a long way off, and attempts to design a simple language for predator agents has similarly failed. Giles and Jim follwed a much more radical approach: They allowed the agents to evolve their own language.
In this experiment, each agent could send a binary message (a short string of ones and zeroes) to a central message board, and could also perceive the messages of the other three agents. With this rudimentary communications framework and no formal language, groups of agents were thrown into random configurations and allowed to pursue their prey for five thousand moves. The programs of the most successful agents were then cross-bred and thrown into new random situations, repeat chorus. In the early phases of the experiment, the messages on the message board were essentially random, but because the agents were designed to try different strategies, and because the ones with the most successful strategies propagated while the least successful died out, they evolved a language over time that allowed them to coordinate their hunt.
More importantly, the researchers found that successful predators evolved language more efficiently if their communication was limited in length in the beginning and grew over time, rather than being uniformly large from the beginning. Expanding the available message size after the predators learned to use shorter "words" allowed the agents to evolve a functioning language much faster. Limiting the message length also seemed to lead the predators to evolve "words" that had different meanings in different situations. (The authors compared this to the word drive, which is a noun or a verb depending on use.) Most astonishing of all, the authors of the study could not always decipher the agents' language. They knew the predators were saying something useful to each other, since they were getting better at chasing down the prey, but finding a Rosetta Stone for human-agentese proved impossible.
The scientific overlaps in this experiment are enormous -- computer science, linguistics, evolutionary theory, even genetics, as the authors encoded the agents' programs into "chromosomes" and shuffled those chromosomes between generations in a process modeled on reproduction. Practical applications for this work are some time off, but its easy to see how Net-crawling agents with evolved language could become part of the Internet's infrastructure.
With Google indexing only a fraction of the Web's content, agents instructing one another about relevant documents could make superior indexing crawlers. With airline tickets fluctuating in price by the minute, traditional purchasing agents have been ineffective because they prevent supply and demand from being measured in one place and real time. A group of agents that can talk to one another, however, can bring the Chicago Solution -- leave the corn in the fields but aggregate the prices in Chicago -- to distributed electronic markets. And with the growth of peer-to-peer networks like Gnutella and Freenet that resist any sort of centralization, such agents may become a necessary accoutrement to making the network user-friendly, or even user-usable.
Computers create problems only computers can solve, and the number of problems where intelligent agents could play a role is multiplying steadily. If Giles and Jim's work turns out to be generally extensible, the question of whether agents will become intelligent enough to help us may take a back seat to whether we are intelligent enough to understand what they are saying to each other.

Clay Shirky is a contributing editor at FEED and Professor of Media Studies at Hunter College.

Other articles by Clay Shirky

Back to Article