Expected Quantization Error Stability-Based Self-Organizing Growth Neural Network With Adaptive Output Network Scale

Self-organizing growth neural network (SOGNN) is an unsupervised clustering algorithm based on competitive learning, which can extract the distribution information and topology information of input data. However, the current SOGNN lacks effective and stable judgment indicators for the scale of network node growth, thereby failing to effectively evaluate and control the reasonable size of the output space. To this end, this paper proposes a judgment index, i.e., expected quantization error stability (EQES), to objectively judge the approximation degree of the output space to the input space. Based on the Growing Neural Gas (GNG) algorithm, an improved GNG algorithm, called GNG-EQES, is proposed, which introduces the EQES criterion to enable the GNG algorithm to generate an appropriate number of output network nodes autonomously without pre-determining the size of the output network. This not only improves the accuracy of feature extraction of the SOGNN, but also improves its adaptive ability and expands its application scope. The experiments in continuous input space and discrete input space have verified the validity and feasibility of the method proposed, and the method is applied to the construction of mobile robot environment topology map.