利用遗传算法改进SOM网络初始权值的乐器分类

针对SOM网络在分类中由于其初始权值的随机性而导致的训练次数过多且易陷入局部最小的问题,提出了利用遗传算法改进网络初始权值的乐器分类。仿真实验提取10种乐器的12阶MFCC系数,之后使用遗传算法计算出每种乐器各阶系数的适应度值,并以此作为网络的初始权值,之后使用已赋初值的SOM网络分类。仿真实验结果表明:利用遗传算法改进SOM网络初始权值的乐器分类方法的分类正确率最高可达到83.51%。

Using Genetic Algorithms to Improve the Initial Weights of SOM Network in the Musical Instrument Classification

For the problem of excessive training and easy to fall into local minimum in SOM network in the classification caused by the randomness of its initial weight, using genetic algorithm to improve network initial weights in instrument classification is proposed. Simulation experiments extract 12-order MFCC coefficients of 10 different kinds of musical instruments. Then use the genetic algorithm to calculate the fitness value of each order in each instrument, and use the fitness value as the network initial weights. Simulation results show that： the way of using genetic algorithms to improve the initial weights of SOM network in the musical instrument classification is effective and the classification accuracy can reach 83.51%.