Akaike Information Criterion-Based Objective for Belief Rule Base Optimization

With the ability in addressing different types of information under uncertainty, the belief rule base (BRB) has been an efficient tool in modeling the nonlinearity of practical systems, as well as taking the experts knowledge and experience into the modeling process. However, the modeling accuracy has been the sole objective for BRB training and learning in the parameter optimization process, which does not take the modeling complexity into consideration. The exclusion of the modeling complexity can directly cause the infeasibility for constructing and further optimizing the BRB system, especially with human's involvement. In this study, the Akaike Information Criterion (AIC) is used the replace the conventional mean square error (MSE) as the modeling objective. Through a thorough deduction process, the AIC-based objective can represent both modeling accuracy and complexity. Furthermore, the BRB parameter model and the corresponding algorithm are proposed as well. The proposed BRB optimization with AIC-based objective is validated by a numeric case study.