This　paper　constructs　an　operant　conditioning　learning　system　based　on　fuzzy　and　probabilistic　automata,　which　used　for　on-line　self-learning　of　fuzzy　rules.　The　learning　system　can　learn　its　rules　on　line　by　interaction　with　environment,　and　achieve　the　best　rule　consequent.　The　probability　cans　grantee　the　global　superiority　of　learning　mechanism.　The　fuzzy　inference　can　improve　the　robustness　and　rapidity　and　of　learning.　Furthermore,　we　adopt　two　fuzzy　control　structure　in　order　to　avoid　rule　explosion　problem　that　mean　the　rules　will　increase　in　exponent　which　induced　by　the　more　number　of　input　variable,　which　can　predigest　difficulty　of　design.　We　apply　our　model　to　inverted　pendulum　self-balance　control　and　the　simulation　indicate:　the　designed　operant　conditioning　learning　system　can　realize　the　self-learning　of　fuzzy　rules,　and　it　especially　has　outstanding　superiority　for　dealing　with　lack　of　prior　knowledge.