The effects of bias, population migration and credit assignment in optimizing trait-based heterogeneous populations

Abstract

Population based search algorithms are becoming the mainstay in nonlinear problems with discontinuous search domains. The generic name of genetic algorithms (GAs) basicly applies to all population based methods. GAs have spawned many versions to suit new applications. Some of these alterations have reached such points that the algorithms may no longer be called GAs. One similar study may be found in [1], in which a perturbation based search algorithm was proposed, called Responsive Perturbation Algorithm (RPA). In a later work [2], instead of a population of homogenous individuals, as is the case for generic GAs, a population of heterogeneous individuals has been set to compete. Replacing the set of winner parents, the fittest individual is made the parent to yield offspring. The current work is now called, with the supplements, trait-based heterogeneous populations plus (TbHP+). Credit assignment and bias concepts in the form of immunity and instinct has been added to provide the populations with a more efficient guidance. Simulations were made through an RBF neural network training, as it was carried out in earlier works, mentioned above, for comparison. Results were prsented at the end as network testing errors which showed further improvement with TbHP+.