First published July 1 2016
Understanding Language Evolution in Overlapping Generations of Reinforcement Learning Agents
Lewys Brace and Seth Bullock
Understanding how the dynamics of language learning and language change are influenced by the population structure of language users is crucial to understanding how lexical items and grammatical rules become established within the context of the cultural evolution of human language. This paper extends the recent body of work on the development of term-based languages through signalling games by exploring signalling game dynamics in a social population with overlapping generations. Specifically, we present a model with a dynamic population of agents, consisting of both mature and immature language users, where the latter learn from the formers interactions with one another before reaching maturity. It is shown that populations in which mature individuals converse with many partners are more able to solve more complex signalling games. While interacting with a higher number of individuals initially makes it more difficult for language users to establish a conventionalised language, doing so leads to increased diversity within the input for language learners, and that this prevents them from developing the more idiosyncratic language that emerge when agents only interact with a small number of individuals. This, in turn, prevents the signalling conventions having to be renegotiated with each new generation of language users, resulting in the emerging language being more stable over subsequent generations of language users. Furthermore, it is shown that allowing the children of language users to interact with one another is beneficial to the communicative success of the population when the number of partners that mature agents interact with is low.