An universal tabu search algorithm for the optimization of functions with continuous variables is presented based on the tabu method so far used. Essentially, the improvements include the transition criterion for different states, the determination of the step vector, the cancellation of tabu list, the stop criteria, and the restart from the optimum etc.. The numerical performances of the present algorithm are investigated using a benchmark problem and the geometry optimization of the multisection arc pole shoe in large salient pole synchronous generators.