当前位置：首页 > news >正文

基于MATLAB实现的PSO优化BP神经网络

news 2025/8/13 9:57:30

基于MATLAB实现的PSO优化BP神经网络的代码，结合了粒子群优化算法（PSO）和BP神经网络，用于提高预测精度。

1. 初始化参数

% 清除环境变量
clear;
clc;% 定义BP神经网络结构
inputLayerSize = 7; % 输入层节点数
hiddenLayerSize = 7; % 隐藏层节点数
outputLayerSize = 1; % 输出层节点数% 定义PSO参数
numParticles = 30; % 粒子数量
maxIter = 100; % 最大迭代次数
c1 = 2.05; % 个体学习因子
c2 = 2.05; % 社会学习因子
w = 0.729; % 惯性权重

2. 读取数据

% 假设数据集为MATLAB自带的脂肪数据集
data = load('bodyfat.mat'); % 加载数据集
inputData = data.data(:, 1:end-1); % 输入数据
outputData = data.data(:, end); % 输出数据% 数据归一化
[inputNorm, inputSettings] = mapminmax(inputData);
[outputNorm, outputSettings] = mapminmax(outputData);% 划分训练集和测试集
[trainInd, testInd] = dividerand(size(inputNorm, 1), 0.7, 0.3);
trainInput = inputNorm(trainInd, :);
trainOutput = outputNorm(trainInd, :);
testInput = inputNorm(testInd, :);
testOutput = outputNorm(testInd, :);

3. 初始化粒子群

% 初始化粒子位置和速度
numDimensions = inputLayerSize * hiddenLayerSize + hiddenLayerSize + hiddenLayerSize * outputLayerSize + outputLayerSize;
particlePosition = rand(numParticles, numDimensions) * 2 - 1; % 随机初始化位置
particleVelocity = zeros(numParticles, numDimensions); % 初始化速度为零
particleBestPosition = particlePosition; % 个体最优位置
particleBestFitness = inf(numParticles, 1); % 个体最优适应度
globalBestFitness = inf; % 全局最优适应度
globalBestPosition = zeros(1, numDimensions); % 全局最优位置

4. 定义适应度函数

% 适应度函数：BP神经网络的均方误差
function fitness = calculateFitness(position, inputLayerSize, hiddenLayerSize, outputLayerSize, trainInput, trainOutput)% 重构BP神经网络的权值和阈值IW = reshape(position(1:inputLayerSize*hiddenLayerSize), [inputLayerSize, hiddenLayerSize]);b1 = position(inputLayerSize*hiddenLayerSize+1:inputLayerSize*hiddenLayerSize+hiddenLayerSize);LW = reshape(position(inputLayerSize*hiddenLayerSize+hiddenLayerSize+1:end-outputLayerSize), [hiddenLayerSize, outputLayerSize]);b2 = position(end-outputLayerSize+1:end);% 构建BP神经网络net = feedforwardnet(hiddenLayerSize);net.IW{1} = IW;net.b{1} = b1;net.LW{2} = LW;net.b{2} = b2;% 训练BP神经网络net = train(net, trainInput', trainOutput');% 计算适应度值output = net(trainInput');fitness = perform(net, trainOutput', output);
end

5. 粒子群优化过程

for iter = 1:maxIterfor i = 1:numParticles% 计算当前粒子的适应度值currentFitness = calculateFitness(particlePosition(i, :), inputLayerSize, hiddenLayerSize, outputLayerSize, trainInput, trainOutput);% 更新个体最优位置和适应度值if currentFitness < particleBestFitness(i)particleBestFitness(i) = currentFitness;particleBestPosition(i, :) = particlePosition(i, :);end% 更新全局最优位置和适应度值if currentFitness < globalBestFitnessglobalBestFitness = currentFitness;globalBestPosition = particlePosition(i, :);endend% 更新粒子速度和位置for i = 1:numParticlesparticleVelocity(i, :) = w * particleVelocity(i, :) + c1 * rand * (particleBestPosition(i, :) - particlePosition(i, :)) + c2 * rand * (globalBestPosition - particlePosition(i, :));particlePosition(i, :) = particlePosition(i, :) + particleVelocity(i, :);end% 显示当前迭代信息fprintf('Iteration %d: Best Fitness = %.6f\n', iter, globalBestFitness);
end

6. 使用优化后的权值和阈值训练BP神经网络

% 重构BP神经网络的权值和阈值
IW = reshape(globalBestPosition(1:inputLayerSize*hiddenLayerSize), [inputLayerSize, hiddenLayerSize]);
b1 = globalBestPosition(inputLayerSize*hiddenLayerSize+1:inputLayerSize*hiddenLayerSize+hiddenLayerSize);
LW = reshape(globalBestPosition(inputLayerSize*hiddenLayerSize+hiddenLayerSize+1:end-outputLayerSize), [hiddenLayerSize, outputLayerSize]);
b2 = globalBestPosition(end-outputLayerSize+1:end);% 构建BP神经网络
net = feedforwardnet(hiddenLayerSize);
net.IW{1} = IW;
net.b{1} = b1;
net.LW{2} = LW;
net.b{2} = b2;% 训练BP神经网络
net = train(net, trainInput', trainOutput');

7. 测试优化后的BP神经网络

% 测试BP神经网络
testOutput = net(testInput');
testOutput = mapminmax('reverse', testOutput, outputSettings);% 计算测试误差
testError = testOutput - mapminmax('reverse', testOutput, outputSettings);
mseError = mse(testError);
fprintf('Test Mean Squared Error: %.6f\n', mseError);