Efficient Sampling with Small Populations

Abstract

We present a population genetic algorithm which satisfies detailed balance, and which has a stationary distribution that factorises into an explicit form for arbitrary fitness functions. For a population size of 1, it is the Metropolis algorithm with a 'breeder' proposal distribution; it extends to larger populations in a natural way, and the stationary (that is, the mutation-selection equilibrium) distribution is exactly known in a simple form for any population size. We term this algorithm exchangeable breeding tuple product sampling (EBT).EBT is closely related to some non-parametric Bayesian Markov-chain Monte Carlo sampling algorithms. EBT can also be viewed as a generalisation of the Moran process.