Numerical experiment with new conjugate direction methods fornondifferentiable optimization

摘要

Results obtained by using a new algorithm for nondifferentiableoptimization are presented. This algorithm is an extension of theWolfe-Lemarechal algorithm, which simulates conjugate direction methodsfor differentiable problems. The author tested two versions of hisalgorithm. The first one is intended for nondifferentiable problemswhose whole subdifferential is easily determined. The second onerequires only one subgradient of the functional at the given point. Theresults reported should be treated qualitatively. The author presentsthe results of the steepest descent method, the Wolfe-Lemarechalalgorithm, and the new algorithm