Scholarly record
COMSOM NEURAL MAPS AS A NEW MACHINE LEARNING TOOL FOR CLUSTERING, FORECASTING AND DATA EVALUATION PROBLEMS
Abstract
The paper presents a new, developed by the author, concept of an artificial neural map, named COMSOM and inspired by classic SOM neural networks. The initial part of the article shows a general idea of COMSOM model and its structure, which includes two layers of artificial neurons: input layer (with typical task of distributing input values to data processing neurons) and basic layer containing usually a great number of processing elements, located inside a rectangular part of an XY plane, and able to move freely in this area during the learning stage. The Y axis of the rectangle represents an output variable (undergoing forecasting or evaluation) in a considered problem. A COMSOM training algorithm, proposed in next part of the paper, is an iterative machine learning procedure (based on special mathematical rules) and leads to both weight vector adaptation and moving neurons positions along X and Y axes. In each learning step, only currently selected, relatively small number of neurons (belonging to a specifically determined neighbourhood like in a classic SOM networks) undergo learning. Finally, nodes equipped with weight vectors recognizing specific groups of objects (input feature vectors) are placed on proper X and Y positions, and can signalize proper value of output (forecasted or evaluated) variable as a value of Y coordinate. Such a model, after completing a training stage, realizes a projection of n-dimensional feature space (in which considered objects, or actually their feature vectors, are placed) into a 2-dimensional part of plane, with special projection of output variable into the Y axis. The model enables recognition and visualization of clusters of objects, with consideration of their output variable values. During the exploitation stage, when a new pattern is delivered on the input of trained COMSOM network, the winning neuron (that best recognizes object being processed) indicates not only the object membership to a given cluster (with visualization in the XY map), but also indicates the predicted output value on the Y axis. Promising results of application of a COMSOM model, in a problem of electricity load forecasting, has been presented in the last part of the paper. Final conclusion and directions of further research have been indicated.
Publication Impact Profile
Publication details
References5
Intelligent Systems in Management. Theory and Practice. (in Polish, ed. Zieliński J.), PWN, Warsaw, 2000.
Kohonen T., Self-organizing Maps, Springer, Berlin, 1995.
Morajda J., Special Cluster Analysis and Basic Feature Estimation with a Modification of Self-Organizing Map. Proc. of the 7th Int. Conf. “Artificial Intelligence and Soft Computing”, Lecture Notes in Artificial Intelligence, Springer, Berlin, 2004.
Moshou D., Artificial Neural Maps. Concepts, Architectures, Applications. VDM Verlag Dr. Müller, Saarbrücken, 2009.
Rutkowski L., Computational Intelligence, Methods and Techniques, Springer, Berlin, 2008.
View or Download full articleAccess options
SWS access login
Login as SWS Scientific CommitteeLogin as SWS Scientific PartnerLogin as SWS AuthorAuthors and approved SWS contributors will read and export their own linked papers after identity matching by SWS profile, email and SGEM GlobalID.
For librarian assistance: [email protected]
Purchase Instant Access
- Article can be downloaded after successful payment.
- Article may be used according to SWS library access terms.
- Article cannot be redistributed.
