function fitness = CostFunction(optVars)
%% 获取数据
vp_train = evalin('base', 'vp_train');
vt_train = evalin('base', 'vt_train');
%% 输入和输出特征个数
inputSize = size(vp_train{1}, 1);
numResponses = size(vt_train{1}, 1);
%% 设置网络结构
opt.layers = [ ...
sequenceInputLayer(inputSize) % 输入层
gruLayer(optVars.NumOfUnits) % GRU层
reluLayer % Relu激活层
fullyConnectedLayer(numResponses) % 回归层
regressionLayer];
%% 设置网络参数
opt.options = trainingOptions('adam', ... % 优化算法Adam
'MaxEpochs', 500, ... % 最大训练次数
'GradientThreshold', 1, ... % 梯度阈值
'InitialLearnRate', optVars.InitialLearnRate, ... % 初始学习率
'LearnRateSchedule', 'piecewise', ... % 学习率调整
'LearnRateDropPeriod', 400, ... % 训练700次后开始调整学习率
'LearnRateDropFactor', 0.1, ... % 学习率调整因子
'L2Regularization', optVars.L2Regularization, ... % 正则化参数
'ExecutionEnvironment', 'auto',... % 训练环境
'Verbose', 0, ... % 关闭优化过程
'Plots', 'none'); % 不画出曲线
%% 数据的参数
num_size = length(vt_train);
%% 交叉验证程序
indices = crossvalind('Kfold', num_size, 5);
for i = 1 : 5
% 获取第i份数据的索引逻辑值
valid_data = (indices == i);
% 取反，获取第i份训练数据的索引逻辑值
train_data = ~valid_data;
% 1份测试，4份训练
pv_train = vp_train(train_data, :);
tv_train = vt_train(train_data, :);
pv_valid = vp_train(valid_data, :);
tv_valid = vt_train(valid_data, :);
% 建立模型
model = trainNetwork(pv_train, tv_train, opt.layers, opt.options);
% 仿真测试
t_sim = predict(model, pv_valid);
% 适应度值
error(i) = sqrt(double(mse(tv_valid, t_sim)));
end
%% 获取适应度
fitness = mean(error);
end
