Restricted Optimal Modeling Method Supervised by Expectation Error

Relation between the possible minimum error (i.e. expectation error) of empirical model and data quality information (e.g., data scale and noise intensity) is analyzed quantitatively, a concept of ldquorestricted optimal modelrdquo is proposed, methods to estimate expectation error are introduced, an idea to optimize model using expectation error is proposed. Based on this idea, an optimal neural network modeling method is proposed. Its availability and superiority is verified by simulation experiment. Furthermore, a new evaluation index of model, namely error average power (EAP), is proposed, which is suitable to evaluate different modeling methods in simulation experiment.