Abstract
Referential games offer a grounded learning environment for neural agents,that accounts for the functional aspects of language. However, they fail toaccount for another fundamental aspect of human language: Because languages aretransmitted from generation to generation, they have to be learnable by newlanguage users, which makes them subject to cultural evolution. Recent work hasshown that incorporating cultural evolution in referential game results inconsiderable improvements in the properties of the languages that emerge in thegame. In this work, we first substantiate this claim with a different data setand a wider array of evaluation metrics. Then, drawing inspiration fromlinguistic theories of human language evolution, we consider a scenario inwhich not only cultural but also genetic evolution is integrated. As our corecontribution, we introduce the Language Transmission Engine, in which culturalevolution of the language is combined with genetic evolution of the agents'architecture. We show that this co-evolution scenario leads to across-the-boardimprovements on all considered metrics. These results stress that culturalevolution is important for language emergence studies, but also the suitabilityof the architecture itself should be considered.