论文部分内容阅读
基于梯度算法和前馈网络所具有的普遍近似性质,提出了一种新的监督型多目标系统化训练机制。在学习过程的实现中,该训练机制一方面能使参数集合选择适当以避免过适应,另一方面能以较少的计算及存储复杂度使网络输出达到所要的精度,保证网络具有满意的可检验性和通用性。新的算法(PTNT)能够在一个过程里同时考虑神经网络训练的几个方面,并且在训练时间和准确度方面也都优于BP算法及其衍生算法。PTNT算法具有类似于LM算法的收敛性,但存储复杂度远远少于LM的一半。文中通过仿真结果证明这种监督训练机制和前馈网络在不同问题环境下的适用性,评价了其有效性。
Based on the general approximation properties of gradient algorithms and feedforward networks, a new supervisory multi-objective systematic training mechanism is proposed. In the process of learning, the training mechanism can make the parameter set choose appropriate to avoid over-adaptation, on the other hand, the network output can reach the required accuracy with less computation and storage complexity, and the network can be satisfied Testability and versatility. The new algorithm (PTNT) can consider several aspects of neural network training in one process at the same time, and also outperforms the BP algorithm and its derivative algorithm in terms of training time and accuracy. The PTNT algorithm has convergence similar to the LM algorithm, but the storage complexity is far less than half of the LM. The simulation results prove the applicability of this supervisory training mechanism and feedforward network under different problems and evaluate its effectiveness.