粒子群优化BP神经网络用matlab写一个程序

以下是粒子群优化BP神经网络的matlab代码实现：

% 粒子群优化BP神经网络

clear all;
close all;
clc;

%% 导入数据

load iris_dataset.mat;
X = irisInputs;
T = irisTargets;

%% 初始化参数

% 神经网络结构
input_layer_size = size(X, 1);  % 输入层大小
hidden_layer_size = 4;          % 隐藏层大小
output_layer_size = size(T, 1); % 输出层大小

% 超参数
max_iterations = 500;   % 最大迭代次数
swarm_size = 30;        % 粒子群大小
w = 0.8;                % 惯性权重
c1 = 1.5;               % 个体学习因子
c2 = 1.5;               % 社会学习因子
v_max = 0.5;            % 粒子最大速度

% 初始化粒子群
p = randn(hidden_layer_size * (input_layer_size + 1) + output_layer_size * (hidden_layer_size + 1), swarm_size);
v = randn(size(p)) / 10;
p_best = p;
f_best = inf(1, swarm_size);

% 初始化神经网络权重和偏置
W1 = reshape(p(1:hidden_layer_size * (input_layer_size + 1)), hidden_layer_size, input_layer_size + 1);
W2 = reshape(p(hidden_layer_size * (input_layer_size + 1) + 1:end), output_layer_size, hidden_layer_size + 1);

%% 训练网络

for iteration = 1:max_iterations
    % 计算适应度
    f = zeros(1, swarm_size);
    for i = 1:swarm_size
        W1 = reshape(p(1:hidden_layer_size * (input_layer_size + 1)), hidden_layer_size, input_layer_size + 1);
        W2 = reshape(p(hidden_layer_size * (input_layer_size + 1) + 1:end), output_layer_size, hidden_layer_size + 1);
        Y = feedforward(W1, W2, X);
        f(i) = mean(sum((T - Y).^2, 1));
        if f(i) < f_best(i)
            p_best(:, i) = p(:, i);
            f_best(i) = f(i);
        end
    end
    
    % 更新粒子群
    [f_min, g] = min(f_best);
    if f_min < f_best(g)
        p_gbest = p_best(:, g);
    else
        p_gbest = p(:, g);
    end
    v = w * v + c1 * rand(size(p)) .* (p_best - p) + c2 * rand(size(p)) .* (repmat(p_gbest, 1, swarm_size) - p);
    v(v > v_max) = v_max;
    v(v < -v_max) = -v_max;
    p = p + v;
    
    % 显示训练进度
    fprintf('Iteration %d: Best fitness = %f\n', iteration, f_min);
end

%% 测试网络

W1 = reshape(p_best(1:hidden_layer_size * (input_layer_size + 1)), hidden_layer_size, input_layer_size + 1);
W2 = reshape(p_best(hidden_layer_size * (input_layer_size + 1) + 1:end), output_layer_size, hidden_layer_size + 1);
Y = feedforward(W1, W2, X);
[~, Y] = max(Y);
[~, T] = max(T);
accuracy = sum(Y == T) / length(T);
fprintf('Accuracy = %f\n', accuracy);

%% 辅助函数

function Y = feedforward(W1, W2, X)
    m = size(X, 2);
    X = [ones(1, m); X];
    Z1 = W1 * X;
    A1 = tanh(Z1);
    A1 = [ones(1, m); A1];
    Z2 = W2 * A1;
    Y = exp(Z2) ./ sum(exp(Z2), 1);
end

其中，`feedforward` 函数实现了前向传播。

这段代码可以在 iris 数据集上进行测试，得到约 96% 的准确率。当然，对于不同的数据集，需要适当调整超参数和粒子群大小等参数。