The　model　parameters　of　a　classifier　directly　affect　the　classification　results.　According　to　the　traits　of　additional　irrelevant　samples　in　the　learning　process　of　Universum　SVM,　this　paper　optimizes　parameters　with　particle　swarm　optimization　(PSO)　due　to　its　simple　concept,　high　computational　efficiency,　and　less　impact　by　the　changes　of　the　problem　dimension;　therefore,　several　parameters　can　be　simultaneously　optimized.　Besides,　selection　for　fitness　function　is　a　key　factor　in　PSO　algorithm.　According　to　its　unbiased　estimation,　k-fold　cross　validation　error　is　considered　as　the　fitness　value,　by　which　an　evaluation　on　the　particle　can　be　obtained.　Finally,　through　experiment　on　tongue　samples,　the　recognition　accuracy　rates　on　test　samples　before　and　after　optimizing　the　parameters　are　compared.　Result　verifies　the　effectiveness　of　the　proposed　algorithm.