Evolving fuzzy neural hydrocarbon networks: A model based on organic compounds

This paper presents a new evolving intelligent model capable of combining the techniques and concepts of artificial neural networks, fuzzy systems and artificial hydrocarbon networks, in which the latter aggregates concepts of organic chemistry to carry out the training of intelligent models. The proposed model has three layers where the first two form a fuzzy inference system and the third layer is responsible for the defuzzification process through concepts based on the bond between carbons and hydrogens. The fuzzification process of the model is based on the techniques of an autonomous data partitioning algorithm that can elicit the number and centers of the clouds that make up the fuzzy neurons in the first layer of the model. Thereby, an evolving algorithm is employed, which uses the data set as a stream in a single-pass incremental mode (allowing fast processing). This is achieved in an unsupervised manner, and thus, to eliminate possible overfitting problems in the subsequent supervised training process, a Bayesian pruning technique is used to identify the neurons that are finally most relevant to the actual supervised approximation and/or classification problem. To validate the proposed approach, binary pattern classification tests, multi-class classification problems and regression problems were performed. The results obtained and compared with other intelligent models in the literature prove that the approach becomes a model capable of extracting knowledge from data sets and using concepts of organic chemistry to perform learning tasks with high degree of reliability. •We present an ensemble method that is able to extract knowledge from data.•The ensemble method is also efficient as a learning model.•Knowledge extraction is automatically represented as fuzzy rules.•The ensemble method can be adapted in presence of new data.