Probabilistic Neuro-Fuzzy System Based on Greed Partitioning Input Space and It's Fast Combined Learning

The paper introduces a novel probabilistic neuro-fuzzy system, employing grid partitioning of the input space, to tackle pattern classification-recognition tasks in batch and online modes, particularly in scenarios with intersecting classes. This system accommodates patterns that may belong to multiple classes simultaneously, each with varying degrees of fuzzy membership. A distinguishing aspect is its notably reduced number of activation membership functions compared to conventional probability neural networks and fuzzy counterparts. Leveraging the active learning principle, the system adapts by forming new membership functions for labeled data, while unlabeled data facilitates the adjustment of membership function centers using T. Kohonen's self-learning rule. Through rigorous computer experiments, the system demonstrates exceptional learning speed and classification accuracy, underscoring its efficacy in handling complex classification tasks.