Simulating　the　hierarchical　principle　of　human　cognizance　process,　a　hierarchical　reduction　algorithm　of　rough　set　theory　is　proposed.　In　the　algorithm,　the　attributes　are　firstly　partitioned　into　different　parts　allocated　at　several　layers.　Then　the　knowledge　in　the　information　systems　or　decision　systems　is　presented　hierarchically　with　multiple　granularities　at　multiple　layers,　and　the　reduction　is　hierarchically　applied　to　parts　of　attributes　allocated　at　each　layer.　The　algorithm　has　a　strong　applicability　and　a　good　dynamic　property,　and　it　can　be　implemented　in　parallel.　The　foundation　of　the　algorithm　is　justified　on　the　basis　of　information　theory.　The　hierarchical　reduction　algorithm　is　used　to　acquire　the　control　decision　of　a　cement　kiln,　which　verifies　the　effectiveness　of　the　algorithm.