The Emergence of Language Among Autonomous Agents

Agents 2000 conference last month in Barcelona, we went souvenir shopping. On the outskirts of the city, an old man in a dilapidated store had some interesting wares, but he spoke only Catalan and we spoke only English and Polish. Nevertheless, even without a common language, we managed to reach an agreement. We left not only with the souvenirs but also knowing their names in Catalan. How did the purchase proceed? As you might imagine, we pointed to the items we were interested in and the shopkeeper pointed to the appropriate coins in our hands. We even engaged in some bartering by offering fewer coins. Why did the purchase succeed? First, we all shared a common knowledge of buying and selling. Second, this knowledge included the value of money (the medium of exchange). And fundamentally , we all understood each other's basic goals for the transaction. Foreign Agents Now, suppose our autonomous shop-bot agents had been representing us by dealing with the vendor's pricebot, and suppose they didn't share an agent communication language (ACL). What should they know at a fundamental level, what could each point to, and how could they establish a common language? Recent research at the University of Texas at Arlington has shown that agents first establish a common vocabulary, progress to a primitive language similar to human pidgin, then enrich the language's grammar to develop a creole, and eventually arrive at a full-blown ACL. During this process, the vocabulary and grammatical structures most important to the agents' task at hand appear first. Thus, shopbots and pricebots will first learn to communicate about various types of goods and money, while softbots that deal with, say, stock market investing will likely develop a different language. However, we must make some assumptions about the agents. First, the agents must be knowledge based. This means they must have a means to represent facts about the world, expressed as sentences in some (hopefully well-defined) knowledge representation language (KRL). This is a reasonable expectation, because a KRL is most likely to be isomorphic to the real world, and distinct things in the world would have distinct representations in a KRL. The agents' knowledge bases are therefore assumed to consist of concepts (classes) organized into a superclass/sub-class hierarchy, with the leaves of the hierarchy occupied by objects (class instances) identified in the agents' environment. Second, the agents must be purposeful , with well-defined …