论文部分内容阅读
连续过程神经元网络在权函数正交基展开时,基函数个数无法有效确定,因此逼近精度不高.针对该问题,提出一种离散过程神经元网络,使用三次样条数值积分处理离散样本和权值的时域聚合运算.模型训练采用双链量子粒子群完成输入权值的全局寻优,通过量子旋转门和非门完成种群进化.局部使用极限学习,通过Moore-Penrose广义逆计算输出权值.以时间序列预测为例进行仿真实验,结果验证了模型的有效性,且训练收敛能力和逼近能力都有一定程度的提高.
In the continuous process of neuronal networks, the number of basis functions can not be effectively determined when the weight function is orthonormal basis, so the approximation accuracy is not high.Aiming at this problem, a discrete process neural network is proposed, which uses cubic spline numerical integration to process discrete samples And weighted sum of weights.Model training uses double-chain quantum particle swarm to perform global optimization of input weights and completes population evolution by using quantum revolving gates and NOT gates.Local learning using limit learning and Moore-Penrose generalized inverse computation We use the time series prediction as an example to simulate the experiment, the results verify the validity of the model, and the training convergence ability and approximation ability have a certain degree of improvement.