论文部分内容阅读
提出采用周期函数取代单调上升函数作为激活函数,并在此基础上,为了加快网络收敛速度及在学习过程中自适应地得到网络的最佳拓扑结构,提出了基于广义卡尔曼滤波的自适应学习和删剪学习算法,并把上述算法应用于4点异或逻辑分类和时间序列预测中。计算机模拟结果显示,使用该周期激活函数,一个无隐藏层的二层感知机就能够解决异或逻辑关系问题。此外,提出的广义卡尔曼滤波学习及删剪算法不仅能够加快网络的收敛速度,而且能够在学习过程中自适应地优化网络的拓扑结构。
In order to speed up the speed of network convergence and get the best topology of the network adaptively in the learning process, a new adaptive learning based on generalized Kalman filter is proposed, which uses periodic function instead of monotonic ascending function as activation function. And delete learning algorithm, and the above algorithm is applied to 4-point XOR logic classification and time series prediction. Computer simulation results show that using this periodic activation function, a layer-2 sensor without hidden layer can solve the problem of XOR relationship. In addition, the proposed generalized Kalman filter learning and pruning algorithm not only can speed up the convergence of the network, but also can adaptively optimize the network topology in the learning process.