Sender–receiver games, first introduced by David Lewis ([1969]), have received increased attention in recent years as a formal model for the emergence of communication. Skyrms ([2010]) showed that simple models of reinforcement learning often succeed in forming efficient, albeit not necessarily minimal, signalling systems for a large family of games. Later, Alexander et al. ([2012]) showed that reinforcement learning, combined with forgetting, frequently produced both efficient and minimal signalling systems. In this article, I define a ‘dynamic’ sender–receiver game in which (...) the state–action pairs are not held constant over time and show that neither of these two models of learning learn to signal in this environment. However, a model of reinforcement learning with discounting of the past does learn to signal; it also gives rise to the phenomenon of linguistic drift. 1 Introduction2 Dynamic Signalling Games with Reinforcement Learning2.1 Introducing new states2.2 Swapping state–action pairs3 Discounting the Past3.1 Learning to signal in a dynamic world3.2 An unexpected outcome: linguistic drift4 ConclusionAppendix: A Markov Chain Analysis. (shrink)

Amodel for inventing newsignals is introduced in the context of sender–receiver games with reinforcement learning. If the invention parameter is set to zero, it reduces to basic Roth–Erev learning applied to acts rather than strategies, as in Argiento et al. (Stoch. Process. Appl. 119:373–390, 2009). If every act is uniformly reinforced in every state it reduces to the Chinese Restaurant Process—also known as the Hoppe–Pólya urn—applied to each act. The dynamics can move players from one signaling game to another during (...) the learning process. Invention helps agents avoid pooling and partial pooling equilibria. (shrink)