Worst-case parameter estimation with bounded model uncertainties