Abstract
The CMAC has been paid attention as a neurological model in the hope of achieving human-like performance in a nonlinear controller design. The CMAC can approximate a wide variety of nonlinear functions by training. The training speed is very fast since the computation is local. Therefore CMAC is well applicable to a large scale system. On the other hand, the disadvantage is that CMAC requires an extremely large memory capacity since it is based on a table look-up method. Furthermore, a trade off between a memory size and an approximation accuracy is required in the choice of the quantization interval.
In this paper, we design a CMAC that has a hierarchal memory structure. The memory structure enables a dynamical segmentation of memory cells. In the region where the output of the nonlinear function to be learned changes rapidly, the quantization interval automatically gets smaller so that the approximation accuracy is improved. This technique ensures an efficient memory usage and alleviates a difficulty of the choice of quantization interval.
Taking a learning control of two degrees robot manipulator as an example, computer simulations are performed to show the validity of the proposed CMAC. The simulation results show that the CMAC requires 1/3 less memory than the previous one without degrading the control performance.
