Advantages of the Growing Cell Structure

The previous section covered the mechanics of the GCS. Since the GCS is used in applications for which we might otherwise use the SOM, we shall briefly consider the advantages of the GCS and the conditions under which the GCS can compete effectively with, or surpass the performance of, SOMs. These include  

The finished network automatically reflects the characteristics of the data domain Not only do the network weights evolve so that they describe the data as fully as possible, but so also does the network geometry. The size of the network is not chosen in advance and as topology is determined by the algorithm and the dataset in combination, it is more likely to be appropriate than the geometry used for a SOM, especially in the hands of an inexperienced user, who might find it difficult to choose an appropriate size of network or suitable values for the adjustable parameters in the SOM. 

Fewer parameters are needed to define evolution of the map In particular, decay schedule parameters are not needed because the size and the shape of the neighborhood vary as the algorithm runs. Nor is it necessary to decide what form of neighborhood function to use, as all neighbors of the BMU are in an identical position. 

It is possible to interrupt a learning process part way through if more data become available or if some data are revised The GCS is, therefore, well suited to the development of incremental or dynamic learning systems. 

The GCS should generate a map of minimum size While the number of nodes on an n x n SOM is unavoidably the square of an integer value and thus may contain more nodes that are actually required for the development of a fully optimized map, the number of nodes in a GCS is not restricted in this way, thus the smallest and, therefore, most economical, map that describes the data can be constructed. 

---