摘 要: 遗传算法作为一种模拟生物遗传进化过程的随机搜索算法,具有并行和全局搜索能力、不要求函数可导等特点,在神经网络学习中得到广泛应用.合理选择初始群体和控制搜索的盲目性,有利于提高算法的效率.为此,提出了一种新的神经网络学习算法———基于样本划分的启发式遗传BP算法.该方法对神经网络学习样本进行划分,形成样本子集.初始群体通过在样本集类上训练神经网络而获得.这些初始网络中包含了关于样本子集的有用信息,根据模式定理,能通过遗传算法保留和加强.此外,提出并证明了关于样本集类及其BP训练的几个代数性质,结合子空间划分进行启发式搜索,以克服搜索的盲目性.对上述方法进行仿真实验,迭代次数和误差较小,表明本学习算法是切实可行的.
（College of Information Science and Engineering, Central South University, Changsha 410083, China）
Abstract:As a random search method mimicking heredity and evolution of creature, genetic algorithm has an ability of parallel and global search with out the function derivable. It has been widely used in neural network learning. To get high efficiency of the algorithm, the initial population should be chosen properly and the blindness of search should be controlled. In this paper, a novel neural network learning algorithm, sample division based heuristic genetic BP algorithm was proposed. In this algorithm, by dividing the samples into subsets, the powerset was formed and the initial population was obtained by training neural networks with BP on the powerset. Using schema theory, useful information about samples, which may be contained in those initial neural networks, can be strengthened and remained. Several algebraic properties of sample sets and their BP training are proposed and proved. Heuristic search is also performed around divided subspaces. The result of simulation shows the validation of this algorithm.
Key words: neural networks; sample; learning algorithm; genetic algorithm