Language Evolution on a Dynamic Social Network

We study the role of the interaction network in a collaborative learning model known as the classification game. This involves a population of agents collaboratively learning to solve a classification task. We have previously shown that interaction while learning allows the agents to converge upon a simpler solution, which in turns guarantees better generalization. This clearly shows the benefits of social learning over individual learning. Here we show that by varying the interaction topology during the learning process, convergence can be achieved very rapidly, at the expense of only a small decrease in the quality of the solution. The interaction topology is evolved using Noisy Preferential Attachment (NPA) algorithm, which is a variant of preferential attachment that unifies the network dynamics with evolutionary dynamics. In the classification game, NPA can be viewed as a recommendation mechanism. We also compare convergence times for several fixed topologies, and show that convergence using NPA is the fastest.