Inoculation to initialise evolutionary search

Abstract

An important factor in the successful application of evolutionary techniques to real-world problems is the incorporation of domain knowledge. One form such knowledge often takes is the possession of one or more high-quality solutions. Non-random initialisation, or inoculation, of the population in an evolutionary algorithm provides a way to incorporate such knowledge. A body of folklore about the methods and results of such initialisation techniques exists, but is largely unwritten and unquantified. This paper discusses the need for hybridisation, through whatever means, and concentrates on the potential offered by seeding the initial population with extant good solutions. Such ideas also have implications for algorithmic restarts after convergence. Experiments conducted using a number of real industrial and commercial problems confirm some of the accepted folklore, and highlight several interesting new results. In particular, it is found that both average solution quality and run-times improve when reasonable inoculation strategies are used, but that the quality of the best solution found over a number of runs often deteriorates as the initial populations become less random.