With　the　increasing　demand　of　wireless　communication,　the　data　rates　supported　by　wireless　communication　technology　are　required　to　be　higher,　which　also　leads　to　the　shortage　of　spectrum　resources.　Cognitive　radio　(CR)　is　considered　a　promising　technology　that　can　make　full　use　of　spectrum　resources.　Currently,　Intelligent　Reflecting　Surfaces　(IRSs)　are　gaining　popularity.　They　autonomously　regulate　the　wireless　communication　environment　to　achieve　higher　data　rates　and　better　coverage.　In　this　paper,　an　IRS-based　model　is　proposed　to　solve　the　energy　efficiency　(EE)　optimization　problem,　which　is　in　multiuser　multiple　input　single　output　(MISO)　energy　harvesting　CR　(EH-CR)　communication　systems.　The　model　enables　the　secondary　transmitter　(ST)　to　utilize　EH　technology　to　capture　radio　frequency　(RF)　energy　from　the　primary　user　(PU),　thereby　powering　its　own　data　transmission.　Due　to　the　complexity　and　continuity　of　action　and　state　space,　a　softmax　deep　double　deterministic　policy　gradients　method　based　on　prioritized　experience　replay　(PER-SD3)　is　designed　to　jointly　optimize　the　transmission　beamforming　vector　of　the　ST　and　the　IRS　reflection　phase　shift　matrix.　Simulation　results　show　that　the　proposed　method　can　improve　the　EE　of　underlay　EH-CR　communication　systems　by　35.3%　at　most　compared　with　the　benchmark　cases.　　©　2015　IEEE.