An adaptive control based on the combination of a novel branch of Soft Computing and fractional order derivatives was applied to control two incompletely modeled, nonlinear, coupled dynamic systems. Each of them contained one internal degree of freedom neither directly modeled/observed nor actuated. As alternatives the decentralized and the centralized control approaches were considered. In each case, as a starting point, a simple, incomplete dynamic model predicting the state-propagation of the modeled axes was applied. In the centralized approach this model contained all the observable and controllable joints. In the decentralized approach two similar initial models were applied for the two coupled subsystems separately. The controllers were restricted to the observation of the generalized coordinates modeled by them. It was expected that both approaches had to be efficient and successful. Simulation examples are resented for the control of two double pendulum-cart systems coupled by a spring and two bumpers modeled by a quasi-singular potential. It was found that both approaches were able to "learn" and to manage this control task with a very similar efficiency. In both cases the application of near integer order derivatives means serious factor of stabilization and elimination of undesirable fluctuations. Since in many technical fields the application of simple decentralized controllers is desirable the present approach seems to be promising and deserves further attention and research.