Regression testing is an important part of the iterative software development, and test data generation is the premise of regression testing. Traditional regression testing methods select a part of test data from existing ones, and generate a number of new test data, so as to verify the correctness of a program. However, these methods are prone to generate redundant test data, therefore reducing the efficiency of regression testing. This paper researches the problem of branch coverage for regression testing, makes full use of information on the path coverage from existing test data, and selects a certain number of paths to cover all the target branches. First, taking a set consisting of several paths as the decision variable, the least number of paths, the largest number of covered branches, and the least number of uncovered paths as the objectives, a 3-objective optimization model for the problem of selecting paths is established. Then, a strategy of evaluating an individual based on the importance of an objective is designed in solving the above model using genetic algorithm. Finally, test data required to generate is determined according to the relation between existing test data and the selected paths. The proposed method is applied to test six benchmark and industrial programs, and compared with other regression testing methods. The experimental results show that paths selected by the proposed method can cover more branches, with fewer test data required to generate and less time consumption for regression testing.