Computational experience with penalty-barrier methods for nonlinear programming

It was recently shown that modified barrier methods are not only theoretically but also computationally superior to classic barrier methods when applied to general nonlinear problems. In this paper, a penalty-barrier function is presented that was designed to overcome particular problems associated with modified log-barrier functions. A quadratic extrapolation of logarithmic terms as well as handling simple bounds separately are utilized. The resulting penalty-barrier method is outlined and compared to previous methods. The conclusion drawn from the computational tests is that the revised method exhibits superior performance on the test set of this study and consequently holds promise as a viable technique for general nonlinear programming.