Towards a Statistical Analysis of Emergent Communication in Evolved Virtual Agents

0 Summary/Abstract Summary/Abstract The field of Evolutionary Robotics has given Cognitive Scientists a new way to study Cognition. Using simulated, embodied agents one can seek to understand the role of basic mechanisms in cognitive processes and how neural control structures give rise to these processes. One such process is language whose minimal cognitive correlate, communication, has been studied by de Greeff and Nolfi in an experiment in which they successfully evolved communicative abilities in virtual agents. While demonstrating that communication can be evolved, questions remain as to exactly how this occurs given that a communicative act requires both the production of a meaningful signal and the appropriate reaction to this signal. Since it is unlikely that these two abilities emerge at the same point, it has been proposed that either a relevant but purposeless production or receptive sensitivity arises first and is then either exploited or exapted to give rise to functional communication. With this work I wish to show that a statistical approach, unlike the original behavioral analysis, can provide certain answers or at least clues regarding the emergence of these abilities. Taking a specific evolved agent and developing information-theoretic measures I demonstrate the utility of these measures in providing evidence for the presence of a producer bias. I then try to generalize these results to other evolved agents. I conclude that while this approach appears promising there is still work to do to create a more generic analysis.