The words we can all agree on: a computer-generated environment to test for evolution of lexical agreement

Abstract

This paper reports on an experiment in which virtual agents in a computer simulated program were set up to assimilate the evolution of lexical agreement based on Complex Adaptive Systems theory (Lee & Schumann 2003, Waldrop, 1992), and Artificial Life (Kirby 2002). Prior studies (Matsen & Nowak 2004; Kirby 2001) have shown that, through iterated learning, computer agents can agree to map a sign to a single referent. In the current experiment, initially agents were randomly assigned to different letter signs. A simple rule based on cellular automata environment (Wolfram 2002) allowed agents to either converge with the sign of their neighbor or change the sign of their neighbor to one similar to theirs. Results of the experiment showed that only under certain conditions could a general consensus on meaning-signal mapping be achieved. If agents had a very loose rule that involved maximum degrees of freedom for agreeing with the neighboring agents, the environment became unstable and no consensus was reached. Similarly, a strict rule with few degrees of freedom produced clusterization with no overall agreement in the environment. The optimal situation for a consensus was achieved under the condition in which agents had more degrees of freedom to agree with a neighboring agent, as well as having less heterogeneity of signs. Furthermore, it is discussed that contrary to biological evolution, patterns of organization such as lexical agreement can emerge out of mere agents’ interaction with one another without the need for the presence of an advantageous trait in the agent as required by Darwinian evolution through natural selection.