In the previous study, the Self-Organizing Nerural-Net-Controller System (SONCS) was developed as an adaptive control system. In order to utilize this system more efficiently, it is important to investigate modelling ability and learning flexibility of the forward model networks. Under the everchanging environmental condition, the forward model should be able to express the complex dynamics of the controlled object and should be easily modified. In this study, the forward model network is improved by adding recurrent connections from the hidden layer to the input layer instead of recurrent connections in the input layer of the previous one. Characteristics of both the forward model networks are investigated using a simple nonlinear system as the modelled object. It is concluded that the frequency range the forward model can cover is spread by this improvement of the recurrent connections. A small test-bed, which has been built to be used to test the control system in the real environment, is introduced.The new version of SONCS is tested using this test-bed through free-swimming tank tests. It is shown that the SONCS can appropriately generate and adjust the controller which is intended to let the vehicle swim at the desired depth.