Abstract: Problems on a probablistic nerual network (PNN) for pattern classification are discussed, and three kinds of method of principal component analysis in the PNN, viz.the eigenvalue decomposition method, orthogonal iterating method and the learning subspace method are studied. These methods can reduce the number or the hiddennodes in the PNN and decrease the testing the for the network. Finally, simulating and measured data are used to verify the performanecs of the presented methods .Experimental results prove that these methods are feasible.