In this paper we shall review the common problems associated with PiecewiseLinear Separation incremental algorithms. This kind of neural models yield poorperformances when dealing with some classification problems, due to theevolving schemes used to construct the resulting networks. So as to avoid thisundesirable behavior we shall propose a modification criterion. It is basedupon the definition of a function which will provide information about thequality of the network growth process during the learning phase. This functionis evaluated periodically as the network structure evolves, and will permit, aswe shall show through exhaustive benchmarks, to considerably improve theperformance(measured in terms of network complexity and generalizationcapabilities) offered by the networks generated by these incremental models.
展开▼
