Co-evolution of language and agents in referential games

Anonymous

Abstract:

Referential games offer a grounded learning environment for neural agents which accounts for the fact that language is functionally used to communicate. However, they do not take into account a second constraint considered to be fundamental for the shape of human language: that it must be learnable by new language learners and thus has to overcome a transmission bottleneck.
In this work, we insert such a bottleneck in a referential game, by introducing a changing population of agents in which new agents learn by playing with more experienced agents. We show that mere cultural transmission results in a substantial improvement in language efficiency and communicative success, measured in convergence speed, degree of structure in the emerged languages and within-population consistency of the language. However, as our core contribution, we show that the optimal situation is to co-evolve language and agents. When we allow the agent population to evolve through genotypical evolution, we achieve across the board improvements on all considered metrics. These results stress that for language emergence studies cultural evolution is important, but also the suitability of the architecture itself should be considered.

Keywords: Referential games, Language Emergence, Language Evolution, Genetic Evolution, Compositionality, Multi-agent Communication, Emergent Linguistic Structure

TL;DR: We enable both the cultural evolution of language and the genetic evolution of agents in a referential game, using a new Language Transmission Engine.