Analysis of using RLS in neural fuzzy systems

In this study, we continue our analysis on the use of RLS in neural fuzzy systems. The recursive least square (RLS) algorithms can have great learning performance for neural fuzzy networks. From our previous work, it can be observed that the advantages of using RLS instead of using BP are not so obvious. For the use of forgetting factor in RLS, the idea is to account for the effects of the change in the premise part. In this study, we have observed that the use of a forgetting factor can still have some advantages when the premise part is fixed. The idea is similar to the used of Widrow-Hoff learning concept in the backpropagation learning algorithm. From our experiments, a strong forgetting factor (smaller value) can let the consequent part trace the error in the learning phase. But the testing error becomes very large. When the system capacity is sufficient, a forgetting factor will improve both in the learning phase and in the testing phase. Finally, the initial value of the covariance matrix is considered. The learning capacity will rise when the initial value increases. But it will increase the error tracing phenomenon in the consequent part too. But it is opposite in a system with less learning capacity.