A Bayesian network is a graphical model that allows the encoding of probabilistic relationships among different measurable features of input data, and proves to be a suitable framework for data modeling when used in conjunction with statistical techniques. Consequently, a Bayesian network is a powerful tool in gaining an understanding about the problem domain, and in handling situations where several data entries are missing. It is an efficient approach to avoiding the over-fitting of data.

This paper presents an input-driven, genetic search-based approach to automatically tuning the representation bias of a supervised inductive learning system. One of the author’s goals is to consider different criteria, such as accuracy, complexity, and task-specific measures, and to incorporate them into an efficient, parallel, search-based wrapper.

The fundamentals of the approach are exposed in the first two sections of the paper. A genetic algorithm (GA)-based wrapper and the operators of selection GA and permutation GA are described in the next two sections. Several experimental results, concluding comments, and suggestions for further extensions are given in the final sections of the paper.

In my opinion, the approach is a sound and promising alternative for improving the efficiency of the feature selection and variable ordering process, leading to more efficient learning strategies for the Bayesian network structure.