针对BP(back propagation)神经网络直接逆向模型精度低、耗时长、易振荡等缺点,提出一种联合改进蚁群算法(IACO)与贝叶斯正则化算法(BR)的BP神经网络逆向建模方法。通过改进蚁群算法,根据搜索阶段设置挥发因子、路径优劣程度更新信息素,并在启发式因子中考虑出发点、终点与各节点的间距等,优化正向模型的权值,提高整体模型精度;之后使用L1/2范数的贝叶斯正则化算法逆向迭代正向模型的输入,达到提高网络稳定性的目的。将本文方法应用于可重构功率放大器中,实验结果表明:相比于直接逆向建模方法和自适应η逆向建模方法,本文方法的建模精度分别提高99.77%、90.70%,平均运行时间分别减少35.76%、2.05%;本文方法可降低功放设计的复杂度,提高其设计速度。

Considering the disadvantages of the direct inverse model for the back propagation (BP) neural network, such as low precision, excessive time consumption, and easy to concussion, this paper proposes an inverse modeling method for the BP neural network that combines an improved ant colony algorithm and a Bayesian regularization algorithm. This method improves the ant colony algorithm, which sets the volatilization factor based on the search stage, updates the pheromone based on the degree of pros and cons of the path, and considers the distance between the starting point and the nodes and the distance between the end point and the nodes in the heuristic factor, to optimize the weight of the forward model and improve the accuracy of the overall model. Then the Bayesian regularization algorithm with L_1/2 norm is used to reverse the input of the forward model, which improves the stability of the network. It is applied to a reconfigurable power amplifier. Experimental results show that the accuracy of the method, compared with that of the direct inverse modeling method and the adaptive η inverse modeling method, is improved by 99.77% and 90.70%, respectively, with the average running time for the modeling being shorten by 35.76% and 2.05%, respectively. Thus, the complexity of designing a power amplification module is reduced and its design speed is accelerated.