AGAPE

AGAPE: Parallel genetic algorithm programming environment developed for APE100/Quadratics. A software environment for the APE100/Quadratics parallel supercomputer is described, aimed for the implementation of parallel evolutionary algorithms. The specification of the evolutionary algorithm is to a great extent constrained by the hardware architecture. An elitist coarse grained stepping stone model is used, where one population is placed at each processor and at regular intervals chromosomes are sent from one population to a neighbouring population. In each generation only best chromosomes from the joined set of parental and offspring chromosomes are used to produce tentative members of the new population. Chromosomes are floating point number vectors. The system can be used for numerical optimization problems. As the source code is freely available, the algorithm can be adapted by user. A version dedicated to a design of various neural networks (feed-forward, recurrent, layered, with supervised and unsupervised learning) is described, with a possibility to learn both weights and topology. A comparison among different evolutionary algorithms to learn neural networks is also reported. The extended introduction gives an overview of the current state of parallel genetic algorithms and evolutionary algorithms for optimization of neural networks.