On the use of ε-most-active constraints in an exact penalty function method for nonlinear optimization