On-line adaptive quantization input space in CMAC neural network

Cerebellar model articulation controller (CMAC) is one kind of neural network that imitates the human cerebellum. It has attractive properties of learning ability and generalization capability. However, the conventional CMAC with equal-size quantization cannot well represent the variation of the target function by finite knots. This paper proposes an online adaptive quantization method that is utilized to adaptively partition the input space of CMAC in accordance with the grey relational analysis. Simulation results on the function approximation show that our method performs better than the conventional one in both the learning speed and the learning precision.