Q-learning　is　a　reinforcement　learning　method　to　solve　Markovian　decision　problems　with　incomplete　information.　The　design　of　reward　function　is　an　important　factor　that　affects　the　learning　results　of　Q-learning.　A　method　to　design　the　reward　function　of　Q-learning　based　on　fuzzy　rules　is　introduced　to　improve　the　performance　of　reinforcement　learning,　and　the　method　is　applied　to　traffic　signal　optimal　control.　According　to　different　traffic　condition,　the　switching　time　and　switching　sequence　of　phase　can　be　adapted.　The　performance　of　the　system　is　evaluated　by　Paramics　microcosmic　traffic　simulation　software.　And　the　results　show　that　the　learning　effect　of　Q-learning　based　on　fuzzy　rules　is　better　than　that　of　conventional　Q-learning　for　traffic　signal　control.