Current Proceedings on Technology

Yazarlar: Erkam Guresen, Gulgun Kayakutlu

Konular:-

Anahtar Kelimeler:Constructive neural network algorithms,Self-organizing maps,K-means clustering

Özet: Artificial neural networks (ANN) have been a powerful data mining tool with no prior data assumptions and non-linear reasoning ability. It has been a challenge to find the optimal network size of ANN model. Majority of the previous research is constructive in nature, aiming the enlargement of a small network by adding new nodes until the desired error level is reached. The constructive algorithms have been generally accepted due to their advantages over the pruning algorithms, regularization and population-based search algorithms. This study aims to evaluate constructive neural network algorithms by clustering them. This evaluation involves analyzing their common properties, common drawbacks and discussing possible future constructive neural network algorithms which will give clear understanding of current literature and possible future research directions. Our evaluation of the constructive algorithms is based on seven criteria: four network structure based and three learning properties based. 20 fundamental constructive algorithms are clustered by self-organizing maps and k-means, and obtained results are evaluated with respect to average error calculation method, topographic product and Dunn's Index.

---