In　prescriptive　fire　protection　designs　of　buildings,　exit　location　is　generally　designed　with　the　constraint　of　the　maximum　travel　distance.　And　the　optimal　exit　location　has　not　been　given　enough　attention　to　in　performance-based　fire　protection　designs　of　buildings.　However,　in　order　to　improve　the　reliability　of　successful　evacuation,　the　optimal　exit　location　for　minimizing　evacuation　time　should　be　studied　with　consideration　of　the　uncertainty　of　initial　occupant　density.　Due　to　the　expensive　computational　cost,　the　surrogate-based　optimization　is　proposed　to　search　the　optimal　exit　location,　which　combines　the　polynomial　chaos　expansion　and　genetic　algorithm.　The　polynomial　chaos　expansion　is　used　to　construct　the　surrogate　model　of　evacuation　time　and　then　genetic　algorithm　is　applied　to　the　polynomial　chaos　expansion　of　evacuation　time　to　search　the　optimal　exit　location.　In　order　to　demonstrate　the　proposed　method,　a　fire　compartment　of　commercial　buildings　with　two　exits　is　presented.　The　results　of　this　case　study　suggest　that　the　optimal　layout　of　two　exits　with　the　same　size　should　be　distributed　symmetrically　and　the　optimal　exit　location　is　different　for　different　probability　levels.　When　the　probability　level　of　successful　evacuation　is　low　with　two　narrow　exits,　the　better　method　is　merging　two　exits　into　be　one　exit　around　the　center　of　the　wall,　which　is　also　consistent　with　that　determined　by　minimizing　the　maximum　travel　distance.　However,　when　the　probability　level　of　successful　evacuation　is　high　with　two　narrow　exits,　the　optimal　location　of　two　exits　should　be　around　two　corners　of　the　wall　in　this　case.　This　proposed　method　can　be　used　to　find　the　suitable　exit　location　under　uncertain　initial　occupant　density,　which　provides　a　reference　for　the　performance-based　design　of　building　exits.　©　2017　The　Author(s).