Multi-agent　systems　form　a　particular　type　of　distributed　artificial　intelligence　systems.　In　this　paper　based　on　basic　reinforcement　learning,　multi-agent　reinforcement　learning　with　special　context　is　proposed.　The　method　is　applied　to　RoboCup　to　learn　coordination　among　agents.　In　experiments,　the　game　field　is　divided　into　different　areas,　and　the　action　choice　is　made　dependent　on　the　area　in　which　the　ball　is　currently　located,　and　spaces　of　state　and　action　decreased,　and　then　the　optimal　joint　policy　is　determined.　Experimental　results　show　the　effectiveness　of　the　proposed　approach.