我正在尝试使用 Encog 库训练神经网络。
数据集（约 7000 个示例）在拆分（分为训练（60%）、交叉验证（20%）和测试（20%））之前进行了线性归一化，如下所示：
Min=-1.000000; Max=1.000000; Average=-0.077008
目标（理想）数据集（也是线性归一化的）如下所示：
Min=0.201540; Max=0.791528; Average=0.477080
我像这样初始化网络：
mNetwork = new BasicNetwork();
mNetwork.addLayer(new BasicLayer(null, false, trainingDataSet.getInputSize()));
mNetwork.addLayer(new BasicLayer(new ActivationSigmoid(), true, numberOfNeurons));
mNetwork.addLayer(new BasicLayer(new ActivationSigmoid(), false, trainingDataSet.getIdealSize()));
mNetwork.getStructure().finalizeStructure();
mNetwork.reset();
我使用ResilientPropagation培训师（也尝试Backpropagation过）：
ResilientPropagation training = new ResilientPropagation(mNetwork, mTrainingDataSet);
for (int i = 0; i < mNumberOfIterations; ++i) {
training.iteration();
result.trainingErrors[i] = mNetwork.calculateError(mTrainingDataSet);
result.validationErrors[i] = mNetwork.calculateError(mValidationDataSet);
System.out.println(String.format("Iteration #%d: error=%.8f", i, training.getError()));
}
training.finishTraining();
在训练过程中，训练者报告的错误一般在减少。完成训练后，我转储重量：
0.04274211002929323,-0.5481902707068103,0.28978635361541294,-0.203635994176051,22965.18656660482,22964.992410871928,22966.23882308963,22966.355722230965,22965.036733143017,22964.894030965166,22966.002332259202,22965.177650526788,22966.009842504238,22965.971560546248,22966.257180159628,22966.234150681423,-21348.311232865744,-21640.843082085466,-21057.13217475862,-21347.52051343582,-21347.988714647887,-21641.161098510198,-21057.27275747668,-21348.784123049118,-21347.719149090022,-21639.773689115867,-21057.095487328377,-21348.269878600076,22800.304816865206,23090.894751729396,22799.39388588725,22799.72408290791,22800.249806096508,22799.19823789763,22799.85510732227,22799.99965531053,22799.574773588192,22799.57945236908,22799.12542315293,22799.523065957797
它们通常要么非常大，要么非常小。对于 sigmoid，它最终会得到收敛到某个数字的预测，例如上面的权重（经过 500 次迭代后得到）给了我：
Min=0.532179; Max=0.532179; Average=0.532179
网络或训练配置似乎有问题。如果我的网络具有低方差，至少它会产生一个在目标范围内的结果。如果它遭受高方差，它将匹配目标。而现在，它完全没有达到目标。
即使预测很差，为什么错误会减少并且变得相当低？有人在我上面的例子中看到明显的错误吗？我对神经网络还不是很有经验。