ABSTRACT
A novel approach toward the adaptive control of robots dynamically interacting with an unmodeled environment and having only approximately known dynamic parameters has been developed on the basis of the principles of the Hamiltonian Mechanics. As the different means of modern Soft Computing technology having a more or less uniform architecture independent of the particular details of the problems to be solved by them, the proposed method also has a uniform structure not strictly tailored to the peculiar properties of the mechanical system to be controlled. However, as special kinds of Artificial Neural Networks (ANN) are fit to solve wide but typical classes of tasks, the proposed method is invented for tackling problems related to the control of mechanical devices in which the dominating non-linear coupling originates from the laws of Classical Mechanics. As ANNs have a plenty of free parameters (connection weights and threshold values) the tuning of which means learning, this mechanical model also contains tunable parameters so offering the possibility of learning. In this paper the model's free parameters, possible constraints imposed on them as well as different tuning strategies are compared to each other on the basis of computer simulations. It is concluded that the method based on the canonical formalism of classical mechanics is promising for gaining different solutions to the problem. However, finding the appropriate tuning rule is far not trivial and a wide area is open for further research from this point of view. The simple tuning strategies here investigated serve as basic paradigms open for further development in the direction of more conventional and better understood methods as Genetic Algorithms or other “Evolutionary Computation” approach.