A Lamarckian Hybrid of Differential Evolution and Conjugate Gradients for Neural Network Training

摘要

The paper describes two schemes that follow the model of Lamarckian evolution and combine differential evolution (DE), which is a population-based stochastic global search method, with the local optimization algorithm of conjugate gradients (CG). In the first, each offspring is fine-tuned by CG before competing with their parents. In the other CG is used to improve both parents and offspring in a manner that is completely seamless for individuals that survive more than one generation. Experiments involved training weights of feed-forward neural networks to solve three synthetic and four real-life problems. In six out of seven cases the DE–CG hybrid, which preserves and uses information on each solution’s local optimization process, outperformed two recent variants of DE.