Abstract—We have used measurements taken on real networkto enhance the performance of our radio networkplanning tool. A distribution learning technique is adopted torealize this challenged task. To ensure better generalizationcapabilities of the learning algorithm, a preprocessing ofdata is required and involves the use of a clustering algorithmthat divides the whole learning space into subspaces.In this paper we apply a new fuzzy clustering algorithmto a prediction tool of a third generation (3G) cellularradio network. Results show that the differences observedbetween simulations and measurements can be considerablydiminished and the generalization capacity is enhancedthanks to the proposed clustering algorithm. This algorithmperforms well than classical k-means algorithm.We can thenpredict with enhanced accuracy new configuration for whichwe don’t have measurements, as long as they are not verydifferent from learned configurations.

Index Terms—Radio Network Prediction, Measurements,Distribution Learning, k-means, Fuzzy Clustering

Cite: Zakaria Nouir, Berna Sayrac, Benoˆıt Fouresti´e, Walid Tabbara and Franc¸oise Brouaye, "Generalization Capabilities Enhancement of a Learning System by Fuzzy Space Clustering," Journal of Communications, vol. 2, no. 6, pp. 30-37, 2007.