Convergência global de um método sem derivadas para otimização irrestrita.
Resumo
Resumo: Apresenta-se um método para minimização irrestrita de uma função F : IRn ! IR duas vezes diferençável cujas derivadas estão indisponíveis. Considera-se para tal, um algoritmo iterativo de região de confiança. Durante as iterações a função objetivo é aproximada por modelos quadráticos através de interpolações polinomiais. São considerados n + 1 pontos interpoladores, os quais de_nem unicamente um polinômio linear. Para se obter modelos quadráticos consideram-se Hessianas como quaisquer matrizes simétricas uniformemente limitadas. De umas iterações para outra, os conjuntos de pontos interpoladores sofrem alterações em no máximo um elemento. Além disso, a cada iterações a função objetivo é avaliada uma única vez. O método proposto possui dois tipos de iterações, região de confiança e alternativa. As do tipo de região de confiança têm como objetivo minimizar o modelo na esperança de que grande parte dessa redução seja herdada pela função objetivo. Já as alternativas visam melhorar a disposi_c~ao dos pontos Interpol adores. Apresenta-se este método de forma algorítmica. Prova-se que se o número de iterações é infinito, se a função objetivo é limitada inferiormente e possui derivadas segundas limitadas, então todo ponto de acumulação da seqüência gerada pelo algoritmo é estacionário. Abstract: We present an unconstrained method for minimization of a function F : IRn ! IR twice di_erentiable whose derivatives are unavailable. For this we consider the iterative algorithm of trust region. During the iterations the objective function is approximated by models using polynomial interpolations with n + 1 points, which de_ne only a linear polynomial. In the case of quadratic models, these are based on previous models, where the Hessian is symmetric, it is lower bounded and it is de_ned in an arbitrary manner. The sets of interpolation points from one iteration to another are changed in at most one element. In addition, at each iteration the objective function is calculated only once. The proposed method has two types of iterations, the trust region iteration and the alternative iteration. The trust region iteration aims to minimize the model in the hope that most of this reduction may be inherited by the objective function. The alternative iterations aims to improve the position of the interpolation points. We present this method in a well-de_ned algorithm. We prove that, if the objective function is bounded below, if the second derivatives is also bounded, and if the number of iterations is in_nite, then every accumulation point of the sequence generated by the algorithm is a stationary point.
Collections
- Teses & Dissertações [9331]