Online updating belief-rule-base using Bayesian estimation

Abstract

The common point of traditional BRB parameters training methods is taking the training process as an optimization problem, which may result in overfitting problem when training data is insufficient or contains strong noises. To solve the problems, we propose a novel method based on Bayesian estimation to update parameters of BRB online. While the optimization methods consider BRB parameters as unknown but determinate values, the Bayesian estimation regards BRB parameters as random variables. Instead of finding single optimal values of parameters, the proposed method is to estimate the posterior distribution of BRB parameters and produce prediction outputs by considering all possible parameters. Since the posterior distribution of BRB parameters cannot be calculated by analytical methods due to the nonlinearity of BRB models, the Sequential Monte Carlo (SMC) sampling technique is adopted to on-line approximate the posterior distribution of BRB parameters. A numerical function and a practical case on pipeline leak detection are studied to verify the performance of proposed algorithm.