P300　brain-computer　interface　(BCI)　is　a　very　promising　technology.　However,　the　time-consuming　calibration　before　every　use　reduces　the　convenience　of　P300　BCI.　In　recent　years,　researchers　are　increasingly　concerning　the　studies　on　calibration-free　P300　BCI.　In　this　study,　we　designed　an　on-line　updating　classifier　for　recognizing　P300　based　on　the　fundament　of　reinforcement　learning　and　developed　a　calibration-free　P300　BCI　system.　The　on-line　updating　classifier　can　be　randomly　initiated　and　quickly　optimized　by　learning　from　the　system　feedback.　The　system　uses　the　latest　classifier　to　recognize　P300,　acquires　the　user＇s　feedback,　and　calls　an　on-line　updating　algorithm　to　optimize　the　classifier.　By　this　system,　any　user　can　start　using　P300　BCI　without　having　to　collect　the　training　data.　To　verify　this　system,　we　recruited　7　volunteers　as　the　subjects　to　participate　in　a　brain-controlled　Chinese　Pinyin　experiment　based　on　this　system.　The　mean　accuracies　of　recognizing　P300,　Pinyin　symbols　and　Chinese　words　in　the　on-line　experiment　were　84.69%　and　80.55%　respectively.　The　results　show　the　effectiveness　of　our　on-line　updating　classifier　and　the　feasibility　of　the　developed　system.　©　2021　IEEE.