A GAd (Genetic Algorithm with Degeneration) is an algorithm that employs genetic algorithms and introduces the idea of genetic damage.
In GAd, the information of a damaged rate is added to each gene,
the genes that have lower effectiveness are inactivated using genetic damage,
degeneration is realized and unnecessary model parameters or rules are reduced.
It is difficult to select proper degeneration speed when GAd is used for structural learning of fuzzy rules, because the proper speed depends on each system to be approximated.
If the degeneration speed is low, estimation error becomes sufficient but the reduction of rules is not enough.
If the degeneration speed is high, the reduction of rules is satisfactory but estimation error becomes rather big.
To solve this problem, we propose to divide the process of structural learning into 3 steps: 
the first step to lower the rule parameter values, the second step to delete the parameters actively, and the final step to minimize the estimation error.
It is shown that the improved GAd is an efficient algorithm for the structural learning of RBF(Radial Basis Function)-fuzzy rule-based systems by applying it to the learning of some function identification problems.