Intransitivity revisited coevolutionary dynamics of numbers games

Relative fitness, or "evaluation by tests" is one of the building blocks of coevolution: the only fitness information available is a comparison with other individuals in a population, therefore they evolve in response to each other, without a global fitness to provide a reference. This can lead to failure, in the form of Red Queen Effect, or cycling. Numbers Games have been studied by several authors as minimal models of intransitivities which could lead to cycling. Here we carry out an analytical study of the dynamics of minimalistic coevolutionary algorithms in the presence of intransitivities, focusing on two-dimensional real-valued numbers games. We show that depending on the characteristics of the problem, the coevolutionary (1+1) hill-climber either makes good progress with constant average speed, or fails, behaving as a random walk. Larger populations fail to bring qualitative changes into those pathological problems, but teacher-learner separation does. Thus this exercise ends up revealing a fundamental difference between single- and separate-population coevolutionary dynamics.