The　echo　state　network　(ESN)　has　been　widely　applied　for　nonlinear　system　modeling.　However,　the　too　large　reservoir　size　of　ESN　will　lead　to　overfitting　problem　and　reduce　generalization　performance.　To　balance　reservoir　size　and　training　performance,　the　multi-objective　sparse　echo　state　network　(MOS-ESN)　is　proposed.　Firstly,　the　ESN　design　problem　is　formulated　as　a　two-objective　optimization　problem,　which　is　solved　by　the　decomposition-based　multi-objective　optimization　algorithm　(MOEA/D).　Secondly,　to　accelerate　algorithm　convergence,　the　local　search　strategy　is　designed,　which　combines　the　l(1)　or　l(0)　norm　regularization　and　coordinate　descent　algorithm,　respectively.　Thirdly,　to　produce　more　solutions　around　the　knee　point,　an　adaptive　weight　vectors　updating　method　is　proposed,　which　is　based　on　decision　maker　interest.　Experimental　results　show　that　the　MOS-ESN　outperforms　other　methods　in　terms　of　network　sparseness　and　prediction　accuracy.