Towards Automating the Evolution of Linguistic Competence in Artificial Agents Using Negotiation

The aim of our research is to understand and automate the mechanisms by which language can emerge among artificial, knowledge-based and rational agents. Our ultimate goal is to design and implement agents that, upon encountering other agent(s) with which they do not share an agent communication language, are able to initiate creation of, and further able to evolve and enrich, a mutually understandable agent communication language (ACL).

First, the agents we are interested in are knowledge-based. This means that they have a representation of facts about the world, expressed as a set of sentences in some (hopefully well defined) knowledge representation language (KRL), for example first order logic, description logic, Classic, KL-One, probabilistic logic, or similar. Our assumption that the agents have a preexisting knowledge base complements much of the related work in artificial life and neural network based approaches (see and references in that volume), genetic algorithms based work, and recent work in AI by Luc Steels.

Second, the agents are purposeful. Usually, this is taken to mean that the agents have well defined goals, i.e., the precise description of states of the world they are to bring about. The possibility that agents may have different goals brings up the notion of self-interested (or selfish) agents, which we allow. We further allow a more expressive representation according to which an individual agent's purpose, or preferences, are expressed in terms of a utility function, as postulated by the utility theory.

Third, the agents are rational. This means that the agents perform actions chosen so as to further their preferences, or goals, given what they know. We follow the operationalization of rationality postulated by decision theory, according to which a rational agent ranks actions in terms of the expected utility of their results, and executes the action with the highest expected utility.

We define communication as the phenomenon of one agent (speaker) producing a signal that, when responded to by another agent (hearer), confers some advantage (or the statistical probability of it) to the speaker. This definition is supported by numerous approaches to study of communication in cognitive science. Simply, the communicative act must be purposeful and beneficial to the speaker, or else a rational speaker would not bother to produce it. Using the the framework of decision theory, a communicative act must lead to an increase of the speaker's assessment of it's own expected utility. This approach allows one to treat communication as action (see Austin's postulate in, since it is defined by its effects on the state of knowledge of hearer and speaker.

In this work, we build on the work by Durfee, Gmytrasiewicz and Noh on values of communicative acts, but we address the issue of language creation and evolution. Given that the ability to communicate can be advantageous, the agents may want to enrich their communicative capabilities if they are insufficient to begin with. Specifically, if two interacting agents do not share a common agent communication language (ACL), they may want to initiate its creation and enrichment to allow mutually beneficial communication. This is the driving force behind evolution of linguistic competence: Improving communication allows the agents to interact more efficiently, and conveys an advantage which we measure as an increase in the agents' expected utilities. This approach is different from one taken by Luc Steels in which agents, playing a ``language game'', are directly rewarded for successful communication, rather than the reward being assessed by the agents based on how communication helps them solve a task at hand. As we mentioned, our employing the the knowledge-base approach further sets our work apart from Steels' work, as well as from related research in artificial life and related fields.

We propose that initiation and enrichment of an agent communication language can be accomplished by the mechanism of negotiation, developed in the fields of economics and game theory, and automated in recent work in artificial intelligence. In proposing negotiation as the main component of our framework we are motivated by the process of language development among humans coming from different linguistic backgrounds that have to interact. Under such circumstances, people were found to create a primitive language, called pidgin, further enrich it to more syntactically sophisticated Creole. In this process, people are frequently said to negotiate among themselves the lexicon and the rules of grammar that become accepted as a part of a shared communication language.