I am training a genetic algorithm for classification and strangely, the training error is consistently HIGHER than the validation and test error. The training and validation set are both small size compared to the test set, because I want to make the algorithm learn from as little data as possible. I would've understood the low size of training set compared to the test set is an issue if the test error was higher than the training, but what I am experiencing now (i.e. training error higher than test error) makes no sense to me. Any idea how this can happen?