Analysis of Accurate Learning in Radial Basis Function Neural Network Using Cosine Similarity on Leaf Recognition

RBF network is a method of artificial neural networks that perform hybrid learning, namely supervised learning and unsupervised learning. Basically, problem found in learning of RBF network that is a difficult to determine the centroid exactly and weight of network the less optimal. Generally, to determine the centroid is done by random and updated to use k-means clustering method. However, selection of initial centroid is less precise will have an effect on the decrease of accurate learning on leaf recognition in RBF network. A method is used to measure the level of similarity between the data proposed in this research, ie cosine similarity. By comparing the distance of similarity between each data, will be generated data the highest level of similarity from the others, so that can be used as a centroid in RBF networks. The result of research RBF network using cosine similarity will be compared a common method used in centroid determination, ie k-means clustering. The results of the test on data showed that leaf recognition with RBF network using cosine similarity method obtained an accuracy level of 79.22%, while the leaf recognition with RBF network using k-means clustering method obtained an accuracy level of 63.91%.