Estimating the ratios of the stationary distribution values for markov chains modeling evolutionary algorithms

  • Authors:

  • Boris Mitavskiy;Chris Cannings

  • Affiliations:

  • School of Medicine, University of Sheffield, Sheffield, UK, S10 2RX. B.Mitavskiy@sheffield.ac.uk and Supported by EPSRC EP/D003/05/1 Amorphous Computing Grant.;School of Medicine, University of Sheffield, Sheffield, UK, S10 2RX. C.Cannings@sheffield.ac.uk

  • Venue:
  • Evolutionary Computation
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

The evolutionary algorithm stochastic process is well-known to be Markovian. These have been under investigation in much of the theoretical evolutionary computing research. When the mutation rate is positive, the Markov chain modeling of an evolutionary algorithm is irreducible and, therefore, has a unique stationary distribution. Rather little is known about the stationary distribution. In fact, the only quantitative facts established so far tell us that the stationary distributions of Markov chains modeling evolutionary algorithms concentrate on uniform populations (i.e., those populations consisting of a repeated copy of the same individual). At the same time, knowing the stationary distribution may provide some information about the expected time it takes for the algorithm to reach a certain solution, assessment of the biases due to recombination and selection, and is of importance in population genetics to assess what is called a “genetic load” (see the introduction for more details). In the recent joint works of the first author, some bounds have been established on the rates at which the stationary distribution concentrates on the uniform populations. The primary tool used in these papers is the “quotient construction” method. It turns out that the quotient construction method can be exploited to derive much more informative bounds on ratios of the stationary distribution values of various subsets of the state space. In fact, some of the bounds obtained in the current work are expressed in terms of the parameters involved in all the three main stages of an evolutionary algorithm: namely, selection, recombination, and mutation.