In　this　paper,　we　propose　a　novel　structure　automatic　change　algorithm　for　neural-network.　It　can　solve　the　problem　that　most　neural-networks　can　not　change　the　structure　online.　This　algorithm　consists　of　two　main　steps:　1)　The　computation　of　the　neural-network　ability　to　judge　whether　need　to　add　nodes　to　the　hidden　layer　or　pruning,　we　use　the　improved　support　vector　machine　(SVM)　to　decide　when　and　where　to　change　the　structure　of　neural-network　hidden　layer　in　this　step;　2)　Adjusting　the　parameter　of　the　neural-network,　this　learning　rule　for　the　neural-network　is　a　novel　approach　based　on　the　modified　back-propagation　(BP).　On　the　basis　of　the　former　methods,　we　propose　a　structure　automatic　changed　neural　network　(SACNN).　Finally,　the　SACNN　is　applied　to　track　the　nonlinear　functions,　the　simulation　results　show　that　the　results　by　this　neural　network　perform　better　than　the　former　growing　cell　structure　(GCS)　neural-network.　©　2008　Springer-Verlag　Berlin　Heidelberg.