A cerebellar model articulation controller (CMAC) neural network is an excellent choice for optimizing the control performance for nonlinear systems. CMAC can be used to generate control inputs and controller parameters, acting as a controller parameter 'tuner'. The most desirable feature of a CMAC is its generalization ability. When CMAC is utilized as a controller parameter 'tuner', the CMAC-based controller facilitates the improvement of control performance for untrained signals. A conventional CMAC achieves higher accuracy by increasing the number of labels for each weight table. However, as the number of labels increases, the requirement for memory increases rapidly and generalization ability decreases. Thus, a CMAC in which the number of labels for each weight table can be decided individually is proposed, allowing layers with more labels to boost accuracy and layers with less labels to improve the generalization ability of the CMAC. Furthermore, hierarchical clustering is employed to determine the labels in each weight table, allowing the input space to be quantized adaptively. Additionally, the output of a CMAC is controller gain, and fictitious reference iterative tuning (FRIT) is employed to adjust the weights of CMAC in an offline manner. In this work, the controller used is a proportional-integral-derivative (PID) controller. Finally, the effectiveness of the proposed method is verified numerically through some simulations and experiments.