Globally Convergent Exact Penalty Algorithms for Constrained Optimization

n this paper we define two classes of algorithms for the solution of constrained problems. The first class is based on a continuously differentiable exact penalty function, with the additional inclusion of a barrier term. The second class is based on a similar modification performed on a continuously differentiable exact augmented Lagrangian function. In connection with these functions, an automatic adjustment rule for the penalty parameter is described, which ensures global convergence, and Newton-type schemes are proposed which ensure an ultimate superlinear convergence rate.