ILClass: Error-driven antecedent learning for evolving Takagi-Sugeno classification systems

The purpose of this research work is to go beyond the traditional classification systems in which the set of recognizable categories is predefined at the conception phase and keeps unchanged during its operation. Motivated by the increasing needs of flexible classifiers that can be continuously adapted to cope with dynamic environments, we propose a new evolving classification system and an incremental learning algorithm called ILClass. The classifier is learned in incremental and lifelong manner and able to learn new classes from few samples. Our approach is based on first-order Takagi-Sugeno (TS) system. The main contribution of this paper consists in proposing a global incremental learning paradigm in which antecedent and consequent are learned in synergy, contrary to the existing approaches where they are learned separately. Output feedback is used in controlled manner to bias antecedent adaptation toward difficult data samples in order to improve system accuracy. Our system is evaluated using different well-known benchmarks, with a special focus on its capacity of learning new classes.