Multi-agent　systems　form　a　particular　type　of　distributed　artificial　intelligence　systems.　As　an　important　character　of　players　in　game,　autonomous　agents’　learning　has　become　the　main　direction　of　researchers.　In　this　paper,　based　on　basic　reinforcement　learning,　multi-agent　reinforcement　learning　with　specific　context　is　proposed.　The　method　is　applied　to　RoboCup　to　learn　coordination　among　agents.　In　the　learning,　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.　This　makes　the　state　space　and　the　action　space　decrease.　After　learning　the　optimal　joint　policy　is　determined.　Comparison　experiment　between　stochastic　policy　and　this　optimal　policy　shows　the　effectiveness　of　our　approach.　©　2005,　Springer-Verlag　Berlin　Heidelberg.