Abstract
Since first introduced, computer simulation has been an increasinglyimportant tool in evolutionary linguistics. Recently, with the development ofdeep learning techniques, research in grounded language learning has alsostarted to focus on facilitating the emergence of compositional languageswithout pre-defined elementary linguistic knowledge. In this work, we explorethe emergence of compositional languages for numeric concepts in multi-agentcommunication systems. We demonstrate that compositional language for encodingnumeric concepts can emerge through iterated learning in populations of deepneural network agents. However, language properties greatly depend on the inputrepresentations given to agents. We found that compositional languages onlyemerge if they require less iterations to be fully learnt than othernon-degenerate languages for agents on a given input representation.