多层前向神经网络的自适应禁忌搜索训练

Applying An Adaptive Strategy-Based Tabu Search to Neural Network

针对BP算法属于局部优化算法的不足,提出了一种新的全局优化算法——自适应禁忌搜索作为前向神经网络的训练算法。该算法通过邻域和候选集的相互配合,动态地调整候选集中分别用于集中性搜索与多样性搜索的元素个数,提高了算法运行的质量和效率。以经典的异或问题(XOR)为例,进行了对比研究。实验结果表明,该算法与BP算法相比明显提高了网络的收敛概率和收敛精度。

Aiming at BP algorithm's drawbacks that it is essentially a local optimization algorithm,a novel and global optimication algorithm,Adaptive Tabu Search,is proposed to train feed-forward neural networks. This algorithm im- proves the quality and efficiency of training neural network by adjusting dynamically the numbers of intensification el- ements and diversification elements in candidate list and by the cooperating of neighborhood and candidate list. Taking the classical XOR problem as an example,a compare investigation is implemented. It shows that this algorithm has obviously superior convergence rate and precision compared to the BP algorithm.