一、优化原理与流程
1. 优化目标
- 参数空间：SVM关键参数为惩罚系数
C和RBF核参数gamma - 目标函数：最大化交叉验证准确率（分类）或最小化均方误差（回归）
2. PSO算法流程
graph TD
A[初始化粒子群] --> B{适应度评估}
B --> C[更新个体最优]
C --> D[更新全局最优]
D --> E{终止条件?}
E -->|是| F[输出最优参数]
E -->|否| B
二、代码实现
1. 数据准备与预处理
% 加载数据（示例使用鸢尾花数据集）
load fisheriris
X = meas'; % 特征矩阵 (4维)
Y = grp2idx(species); % 类别标签% 数据归一化
[X, ps_input] = mapminmax(X', 0, 1);
X = X';% 划分训练集/测试集
cv = cvpartition(Y,'HoldOut',0.3);
X_train = X(cv.training,:);
Y_train = Y(cv.training);
X_test = X(cv.test,:);
Y_test = Y(cv.test);
2. PSO参数设置
% PSO参数
pso_option = struct(...'maxgen', 100, % 最大迭代次数'sizepop', 30, % 粒子数量'c1', 1.5, % 个体学习因子'c2', 1.7, % 群体学习因子'w', 0.7, % 惯性权重'vmax', 0.5, % 速度上限'vmin', -0.5, % 速度下限'pc1', 0.1, % 参数c1变异概率'pc2', 0.1); % 参数c2变异概率% SVM参数范围
param_range = [0.01, 100; % C的范围0.001, 10]; % gamma的范围
3. 适应度函数定义
function fitness = svmFitness(params, X_train, Y_train)% 参数解码C = 10^params(1);gamma = 10^params(2);% 交叉验证参数cmd = sprintf('-v 5 -c %f -g %f', C, gamma);% 训练SVM并返回交叉验证准确率accuracy = svmtrain(Y_train, X_train, cmd);fitness = -accuracy; % 最小化问题
end
4. PSO优化过程
% 初始化粒子群
num_particles = pso_option.sizepop;
dim = 2; % C和gamma两个参数
particles = rand(num_particles, dim) .* (param_range(:,2)' - param_range(:,1)') + param_range(:,1)';
velocities = rand(num_particles, dim) .* (pso_option.vmax - pso_option.vmin) + pso_option.vmin;% 初始化个体最优
pbest = particles;
pbest_fitness = inf(num_particles,1);% 初始化全局最优
[gbest_fitness, gbest_idx] = min(pbest_fitness);
gbest = pbest(gbest_idx,:);% 迭代优化
for iter = 1:pso_option.maxgenfor i = 1:num_particles% 适应度评估current_fitness = svmFitness(particles(i,:), X_train, Y_train);% 更新个体最优if current_fitness < pbest_fitness(i)pbest(i,:) = particles(i,:);pbest_fitness(i) = current_fitness;end% 更新全局最优if current_fitness < gbest_fitnessgbest = particles(i,:);gbest_fitness = current_fitness;endend% 速度更新r1 = rand(num_particles,dim);r2 = rand(num_particles,dim);velocities = pso_option.w*velocities ...+ pso_option.c1*r1.*(pbest - particles) ...+ pso_option.c2*r2.*(gbest - particles);% 速度限制velocities = max(min(velocities, pso_option.vmax), pso_option.vmin);% 位置更新particles = particles + velocities;% 参数范围限制particles(:,1) = max(min(particles(:,1), param_range(1,2)), param_range(1,1));particles(:,2) = max(min(particles(:,2), param_range(2,2)), param_range(2,1));% 自适应变异if rand < pso_option.pc1particles(:,1) = 10^randn(num_particles,1);endif rand < pso_option.pc2particles(:,2) = 10^randn(num_particles,1);end% 显示迭代信息fprintf('Iter %d: Best Fitness=%.4f (C=%.4f, gamma=%.4f)\n',...iter, -gbest_fitness, 10^gbest(1), 10^gbest(2));
end
5. 模型训练与评估
% 使用最优参数训练SVM
best_C = 10^gbest(1);
best_gamma = 10^gbest(2);
cmd = sprintf('-c %f -g %f', best_C, best_gamma);
model = svmtrain(Y_train, X_train, cmd);% 测试集预测
[predict_label, accuracy, dec_values] = svmpredict(Y_test, X_test, model);% 结果输出
fprintf('测试集准确率: %.2f%%
', accuracy(1));
三、改进
1. 自适应参数调整
-
惯性权重动态调整：
w = pso_option.w * (1 - iter/pso_option.maxgen); % 线性递减 -
学习因子自适应：
c1 = pso_option.c1 * (1 - iter/pso_option.maxgen); c2 = pso_option.c2 * (1 + iter/pso_option.maxgen);
2. 多目标优化扩展
% 同时优化准确率和模型复杂度
fitness = -accuracy + 0.1*(sum(W.^2)); % W为支持向量权重
3. 并行计算加速
parfor i = 1:num_particles% 并行计算适应度
end
参考代码 小波神经网络预测代码 www.youwenfan.com/contentcni/64776.html
四、可视化分析
1. 参数收敛曲线
figure;
plot(1:pso_option.maxgen, -gbest_fitness*ones(1,pso_option.maxgen), 'r--');
hold on;
plot(1:pso_option.maxgen, mean(pbest_fitness,1), 'b-.');
legend('全局最优', '平均适应度');
xlabel('迭代次数'); ylabel('适应度值');
title('PSO收敛曲线');
2. 参数分布热力图
figure;
histogram2(pbest(:,1), pbest(:,2), 'DisplayStyle','tile');
xlabel('log10(C)'); ylabel('log10(gamma)');
title('最优参数分布');
colorbar;