Pattern classifier net using information entropy

Summary form only given. The authors propose different representations for the training of a node so as to enhance the power of the PSIN (parallel sequential induction network). A binary output representation with higher order connection makes the PSIN more suitable for the classification problem with arbitrary shapes of boundaries. The unary representation with competitive scheme allows the PSIN to deal more easily with both the multiclass problem and the multiple decision regions problem. The numerical examples demonstrate the superiority of the proposed schemes. Comparisons are made against the backpropagation network and conventional schemes.<>