A multilayer feedforward neural network having N/4 nodes in two hidden layers

In order to reduce the complexity of a single hidden layer multilayer neural network, a new two hidden layer MFNN (THL-MFNN) with a combined structure of a RBFN and MLPs is proposed, and its associated training method is discussed. The proposed THL-MFNN can be easily constructed, and can be efficiently trained by online recursive methods. The performance of the proposed THL-MFNN with P/4+2=18 hidden nodes and 34 weights is equal to that of an optimum Bayesian equalizer using an RBFN with P=64 hidden nodes and 64 weights. The role of each layer in the proposed THL-MFNN is presented by a theoretical approach, and the feasibility of a more reduced structure is given.