Abstract
A structural evolving approach (SEA) based on incremental partitioning learning is proposed in this paper. SEA starts with a simple fuzzy system with one fuzzy rule containing no fuzzy term for the antecedent part. After that, it keeps evolving by adding new fuzzy terms to the selected attribute in the selected partition. A new partitioning technique that locates sufficient splitting points is proposed. Furthermore, a dynamic partition-selection technique that leads to the best tradeoff between accuracy and interpretability is presented. In addition, a rule reduction mechanism that is able to locate rules with low and high impact on the system is developed. Finally, a local linear optimization is used to find the consequent parameters, which means that SEA uses only linear systems to find the antecedents and consequents parameters. Therefore, a simple and high interpretability system is offered. Six data sets are used to validate the performance of SEA with similar works. Despite the low complexity of SEA, the results show that SEA outperforms the existing methods with fewer fuzzy rules and fewer antecedent conditions too.