Removing ‘mind-reading’ from the Iterated Learning Model

The iterated learning model (ILM), in which a language comes about via communication pressures exerted over successive generations of agents, has attracted much attention in recent years. Its importance lies in the focus on cultural emergence as opposed to biological evolution. The ILM simplifies a compositional language as the compression of an object space, motivated by a poverty of stimulus—as not all objects in the space will be encountered by an individual in its lifetime. However, in the original ILM, every agent ‘magically’ has a complete understanding of the surrounding object space, which weakens the relevance to natural language evolution. In this paper, we define each agent’s meaning space as an internal self-organising map, allowing it to remain personal and potentially unique. This strengthens the parallels to real language as the agent’s omniscience and ‘mind-reading’ abilities that feature in the original ILM are removed. Additionally, this improvement motivates the compression of the language through a poverty of memory as well as a poverty of stimulus. Analysis of our new implementation shows maintenance of a compositional (structured) language. The effect of a (previously-implicit) generalisation parameter is also analysed; when each agent is able to generalise over a larger number of objects, a more stable compositional language emerges.