In　relational　databases　and　their　applications,　an　important　issue　is　to　evaluate　a　stream　of　top-N　selection　queries.　For　this　issue,　we　propose　a　new　method　with　learning-based　strategies　and　region　clustering　techniques　in　this　paper.　This　method　uses　a　knowledge　base　to　store　related　information　of　some　past　queries,　groups　the　search　regions　of　the　past　queries　into　larger　regions　and　retrieves　the　tuples　from　the　larger　regions.　To　answer　a　newly　submitted　query,　our　method　tries　to　obtain　most　results　from　the　previously　retrieved　tuples　that　are　still　in　main　memory.　Thus,　this　method　seeks　to　minimize　the　response　time　by　reducing　the　search　regions　or　avoiding　accesses　to　the　underlying　databases.　Extensive　experiments　are　carried　out　to　measure　the　performance　of　this　new　strategy　and　the　results　indicate　that　it　is　significantly　better　than　the　naive　method　for　both　low-dimensional　and　highdimensional　data.　©　2008　IEEE.