Self-balancing　two-wheel　robot　is　a　high　order,　multi-variable,　nonlinear,　strong-coupling　and　absolutely　unstable　system.　A　reinforcement　learning　algorithm　based　on　many　parallel　Cerebellar　Model　Articulation　Controller　(CMAC)　neural　networks　is　proposed　for　the　balance-control　problem　of　self-balancing　two-wheel　robot.　In　the　method,　the　outputs　of　CMAC　are　used　to　approximate　the　Q-functions　of　the　input　state　variables.　The　input　state　variables　are　divided　to　decrease　the　grades　of　quantization.　Therefore,　the　storage　spaces　of　CMAC　are　reduced　effectively,　and　the　learning　rate　and　control　precision　of　Q-algorithm　are　improved.　At　the　same　time,　the　generalization　of　continuous　state　variables　is　realized　too.　The　method　is　applied　to　solve　the　balance　control　problem　of　self-balancing　two-wheel　robot,　and　the　simulation　results　show　its　correctness　and　efficiency.