This paper has designed and realized a stock investment decision system based on gene expression programming. The system isn't used to predict the number of stock index, but to make an investment decision according to the future trend of stock index.Firstly, the gene expression programming based theories, methods and applications both at home and abroad, are summarized in this paper. Considering the gene expression programming is a new computational evolution theory, so it is necessary to describe the origin, development and the background of genetic algorithm, including the thoughts and methods in the development process. Secondly,the three branches of genetic algorithm GA, GP and GEP are analyzed respectively. And we research, analyze and compare the important problems such as the genetic representation, fitness function and genetic operator. Also we analyzed the performance of the algorithm convergence efficiency, and the evolution complexity, and proposed an improved method. I believe GEP is the compromise of GA and GP, it improves the performance and convergent efficiency, but complexity of the solution is reduced. This paper suggests changing the genetic representation which doesn't influent the evolution performance and convergence efficiency, raise a method that make genetic representation multilayered and make gene segment modeled. For that idea the problem of reusing genetic fragments is solved, and the gene expression's ability of solving problems is improved.Moreover, I begin with a classic problem called time series prediction which stock index prediction problem belong to, and have summarized the current research situation at home and abroad,and have discussed and researched some methods of time series prediction problems. Learn about each prediction method for its good aspect, and analyzed bad aspect of the algorithm, and the deficiency in algorithm is improved.Finally, a stock investment decision system based on gene expression programming is designed and realized. It improved original gene expression programming algorithm, produced a multilayer gene representation, solved the reuse problem of the gene string. The karva genetic code method is improved, so the complexity of the algorithm is strengthened, and does not affect the evolution efficiency of algorithm. The fitness function has been improved. It solved the problem of the high fitness but low ability. The genetic operator is improved in order to adapt the new gene representation. In addition, there are some innovations in system design, it does not treat the time series prediction problem as symbolic regression problems, but creates a logical determine method based on stock index trend. This design makes the results having more logic meaning.Experiment results show that the improved algorithm is effective, not only the evolution efficiency and convergence are similar to traditional GEP, but also the complexity of the problem can be handled is greatly improved. |