muARTMAP: use of mutual information for category reduction in Fuzzy ARTMAP.

A new architecture called muARTMAP is proposed to impact a category proliferation problem present in Fuzzy ARTMAP. Under a probabilistic setting, it seeks a partition of the input space that optimizes the mutual information with the output space, but allowing some training error, thus avoiding overfitting. It implements an inter-ART reset mechanism that permits handling exceptions correctly, thus using few categories, especially in high dimensionality problems. It compares favorably to Fuzzy ARTMAP and Boosted ARTMAP in several synthetic benchmarks, being more robust to noise than Fuzzy ARTMAP and degrading less as dimensionality increases. Evaluated on a real-world task, the recognition of handwritten characters, it performs comparably to Fuzzy ARTMAP, while generating a much more compact rule set.

Learning from noisy information in FasArt and FasBack neuro-fuzzy systems.

Neuro-fuzzy systems have been in the focus of recent research as a solution to jointly exploit the main features of fuzzy logic systems and neural networks. Within the application literature, neuro-fuzzy systems can be found as methods for function identification. This approach is supported by theorems that guarantee the possibility of representing arbitrary functions by fuzzy systems. However, due to the fact that real data are often noisy, generation of accurate identifiers is presented as an important problem. Within the Adaptive Resonance Theory (ART), PROBART architecture has been proposed as a solution to this problem. After a detailed comparison of these architectures based on their design principles, the FasArt and FasBack models are proposed. They are neuro-fuzzy identifiers that offer a dual interpretation, as fuzzy logic systems or neural networks. FasArt and FasBack can be trained on noisy data without need of change in their structure or data preprocessing. In the simulation work, a comparative study is carried out on the performances of Fuzzy ARTMAP, PROBART, FasArt and FasBack, focusing on prediction error and network complexity. Results show that FasArt and FasBack clearly enhance the performance of other models in this important problem.