Kitano proposed to use GA with graph encoding method to have good scalability for hierarchical networks. However, this is not applicable to recurrent networks. The authors propose to use a combination of a structural learning with forgetting(SLF) and GA for designing recurrent neural networks; the former generates quasi-optimal recurrent network structure and the latter prevents local minima by global search. Its applications to two kinds of time series data well demonstrate the superiority to SLF and to a combination of GA and BPTT.