An adaptive merging and growing algorithm for designing artificial neural networks

This paper presents a new algorithm, called adaptive merging and growing algorithm (AMGA), for designing artificial neural networks (ANNs). The new algorithm merges and adds hidden neuron during training. The merging operation introduced here is a kind mixed mode operation that is equivalent to pruning two neurons and adding one neuron. Unlike most previous studies on designing ANNs, AMGA puts emphasis on adaptive functioning in designing ANNs. This is the main reason why AMGA merges and adds hidden neurons repeatedly (or alternatively) based on the learning ability of hidden neurons and training progress of ANNs, respectively. AMGA has been tested on five benchmark problems including the Australian credit card, cancer, diabetes, glass and thyroid problems. The experimental results show that AMGA can produce ANNs with good generalization ability compared to other algorithms.