Ellipsometry can not be directly tested by thin film optical parameters, which needs approximate solution method for numerical inversion. Using genetic algorithms, the selection operator, mutation operator and crossover operator are improved by drawing from competition, selection and fitness niche adjustment idea. The improved algorithm is effective in preventing the "premature" phenomenon. It searches the global optimal and reduces the requirements on the operator to set the initial value on the fitted model. In practice, by comparing the results of the three fitting methods, we gain reliable optical constants of diamond-like carbon (DLC) films prepared under certain conditions. The improved genetic algorithm is effective to fit thin film optical constants through a large number of fitting.