In　traditional　cognitive　radio　(CR)　network,　secondary　users　(SUs)　are　always　assumed　to　obey　the　rule　of　＂introducing　no　interference　to　the　primary　users　(PUs)＂.　However,　this　assumption　may　be　not　realistic　as　the　CR　devices　becoming　more　and　more　intelligent　nowadays.　In　this　paper,　with　the　concept　of　light-handed　CR,　which　is　proposed　to　deal　with　the　above　mentioned　problem　by　enforcing　＂　punishment＂　to　illegal　CR　transmissions,　we　model　the　action　decisions　of　primary　users　(PUs)　as　a　partially　observable　Markov　decision　process　(POMDP),　and　propose　the　optimal　spectrum　allocation　scheme　with　the　objective　of　maximizing　their　reward.　Utility　function　is　defined　as　the　reward　in　this　paper　as　well.　Furthermore,　extensive　simulation　results　show　that　the　proposed　scheme　improves　the　reward　significantly　compared　to　the　existing　scheme.