In the first report the Self-Orgasizing Neural-net-Controller System (SONCS) which has been developed by the author's laboratory was applied to generate a controller for constant altitude swimming of AUVs (Autonomous Underwater Vehicles). Although the performance of the resulting network was successfully demonstrated, the adjustment rate of the neural-net controller was considerably slow. In this paper three ideas are proposed to accelerate the adjustment of the controller : 1) Modularization of the forward model network ; 2) Introduction of difference type network ; and 3) Generation of tentative teaching samples for the controller network. The process of the training of a cruising type AUV is demonstrated in computer simulations for verification of effectiveness of these improvements. It is shown that an appropriate controller is quickly constructed after about ten times of training.