Questions regarding the origins of language are notoriously hard. We cannot effectively observe processes of cultural evolution which shape natural languages because of the timescales involved. Hence, computational simulations with populations of artificially evolved agents become useful tools to formalize theories, and to study the consequences of various assumptions regarding language evolution. There are technical and methodological challenges involved, concerning model implementation and interpretation of the results.

Looking for guidelines for good architectures for such simulations, I will analyze examples of existing models. Specifically, I will discuss which learning algorithms are being used, and how they are related to the theoretical assumptions behind the model. Two important groups of techniques will be compared: genetic algorithms, inspired by biological evolution, and reinforcement learning, which may represent individual learning. As I will show, technical motivations for adoption of these techniques within a model may not always be consistent with their customary interpretations. At the end, I will ask which behaviors of agents truly emerge due to the simulated processes of evolution and learning, and which stem directly from their design. This allows to distinguish between intelligence of the agents and knowledge of their constructors. In successful models properties crucial for the tested theory should always emerge on agents' side.