Learning　sample　selection　has　the　very　tremendous　influence　to　the　generalization　ability　of　the　neural　network.　This　paper　proposed　that　the　learning　sample　selection　should　consider　the　characteristics　of　the　non-linear　object　approached.　A　method　of　selecting　learning　samples　based　on　the　uniform　design　and　clustering　theory　is　discussed.　The　result　of　application　shows　that　the　effective　selection　of　the　learning　samples　can　improve　the　generalization　ability　of　the　neural　networks.