A Neural Network Based Hybrid Method For Solving Quadratic Bilevel Programming Problem And Bilevel Portfolio Optimization Model

In this research, we have developed a neural network based hybrid method for solving the quadratic bi- level programming problems. This proposed algorithm combines a genetic algorithm and a recurrent neural network together in order to solve such problems efficiently and accurately. The genetic algorithm is developed for dealing with the upper level problem. It will choose good solution candidates and pass them to the lower level problem. Then, in the lower level, we use the parameterized dual neural network to get possible optimal solutions. The experiment results indicate that compared with other methods, the proposed neural network based hybrid method is capable of achieving better optimal solutions for the quadratic bi- level programming problem in a short time. In the meanwhile, when choosing good initialization parameters, the proposed algorithm is able to obtain the optimal solution with a high accuracy. Since in the research field of bi- level programming problems, there are still few usages of neural network based and hybrid based methods, it is without doubt that this proposed algorithm ca n contribute to the research of such problems. After developing the proposed neural network based hybrid method,we build an application model of quadratic bi- level programming problem, whose name is Bi- Level Portfolio Optimization Model, starting from the consideration of seeking a better approach for modeling the inherent conflict factors in the financial market. Based on the Japan stock market, the experiment results show that the proposed hybrid method can solve this application with a good performance. The solution of the model is one of the optimal point with a reasonable trade-off preference between risk and return. It is evident that the portfolio selections achieved by this model could be a good guidance for the investors during the investment activity.