The　problem　of　over　fitting　in　DBN　is　extensively　focused　on　since　different　networks　may　respond　differently　to　an　unknown　input.　In　this　study,　a　regularization-reinforced　deep　belief　network　(RrDBN)　is　proposed　to　improve　generalization　ability.　In　RrDBN,　a　special　regularization-reinforced　term　is　developed　to　make　the　weights　in　the　unsupervised　training　process　to　attain　a　minimum　magnitude.　Then,　the　non-contributing　weights　are　reduced　and　the　resultant　network　can　represent　the　inter-relations　of　the　input-output　characteristics.　Therefore,　the　optimization　process　is　able　to　obtain　the　minimum-magnitude　weights　of　RrDBN.　Moreover,　contrastive　divergence　is　introduced　to　increase　RrDBN＇s　convergence　speed.　Finally,　RrDBN　is　applied　to　hand-written　numbers　classification　and　water　quality　prediction.　The　results　of　the　experiments　show　that　RrDBN　can　improve　the　recognition　performance　with　less　recognition　errors　than　other　existing　methods.