前向网络的两种混合学习策略

针对前向网络反向传播算法（ＢＰ）训练缓慢和易于陷入局部极小的缺点以及反向运算的复杂性，利用ＢＰ算法监督学习特点、模拟退火（ＳＡ）算法在局部极小处的概率突跳特性和遗传算法（ＧＡ）的并行化群体搜索的特点，有效结合ＢＰ和ＳＡ算法以及ＧＡ和ＳＡ算法，提出了前向网络的两种混合学习策略即ＢＰ＆ＳＡ混合策略和ＧＡ＆ＳＡ混合策略。以异或问题为例，通过计算机仿真对混合策略与ＢＰ、改进ＢＰ算法的比较表明混合学习策略较大程度改进了前向网络学习的收敛性能和收敛速度，并一定程度上避免了反向运算的复杂性，是前向神经网络学习的有效算法。

Two hybrid learning strategies of feedforward network

As back propagation learning algorithm of feedforward network which has large computation complexity is prone to converge slowly and stick into local minima, two hybrid learning strategies of feedforward network, named BP&SA hybrid strategy and GA&SA hybrid strategy, which combine the supervising learning property of BP with the probabilistic jumping property of simulated annealing (SA) algorithm and combine the paralell searching property of genetic algorithms (GA) with SA are presented. As far as exclusive OR problem is concerned, computer simulation results by comparing with BP and improved BP algorithms show that the hybrid strategies are feasible and effective learning algorithms which greatly improve the convergence property and avoid large computation complexity to some extent.