Training recurrent neural networks for dynamic system identification using parallel tabu search algorithm

There are several modern heuristic optimisation techniques, such as neural networks, genetic algorithms, simulated annealing and tabu search algorithms. Of these algorithms, the tabu search is quite a new, promising search technique for numeric problems, especially for nonlinear problems. However, the convergence speed of the standard tabu search to the global optimum is initial-solution-dependent, since it is a form of iterative search. In this paper, a new model of tabu searching, which has been proposed by the authors to overcome the drawback of a standard tabu search, is tested for training a recurrent neural network to identify dynamic systems.