Classical statistical analysis is split into two steps: model selection and post-selection inference and the uncertainty introduced by model selection is often ignored. Model averaging addresses this issue by averaging over a set of candidate models and incorporates the model uncertainty. The study provided a general framework of model averaging procedure based on a weak local misspecification condition that the estimator from candidate model is within o(n(1/4)) neighborhood of true model. The study also employs the optimal weight strategy to provide data driven weights that minimize the asymptotic mean squared error of the estimator. The framework is employed to both generalized linear model, partially linear single-index model and generalized partially linear single index model and is applied in both simulation studies and real data examples. In addition, the study discusses future research area that can be extended from this method.