Enhance the performance of CMAC neural network via fuzzy theory and credit apportionment

Cerebellar model articulation controller (CMAC) is one kind of neural network that imitates the structure of human cerebellum, storing information in different layers. For an all learning process, the disadvantage of conventional CMAC with a larger fixed learning rate is the unstable phenomenon; at the same time, the smaller learning rate will cause the slower convergence speed. In this aspect, we propose a dynamic adjusting learning rate via different situations. Hence, we adopt the fuzzy rule to give an appropriate learning rate to achieve a better response than the conventional CMAC. In addition, in order to speed up the learning speed and reduce the phenomenon of learning interference, we adopt the concept of credit apportionment, giving different credits to different weights depending on their relationships with adjacent states. Simulation result shows that the modified CMAC has a more satisfactory performance than the conventional CMAC.