Abstract: Traditional methods adopt the two-stage learning mechanism of training before testing, which can easily lead to over fitting or under fitting. To solve this problem, an optimization learning algorithm based on hybrid bilevel self-organizing radial basis function neural network (Hb-SRBFNN-OL) was proposed. First, the training process and testing process were integrated into a unified framework to effectively balance the overfitting and underfitting problems. Second, an interactive learning algorithm with two layers was proposed. The upper layer organized and adjusted the network structure based on the network complexity and test error, and the lower layer used Levenberg Marquardt (LM) algorithm as the optimizer to optimize the connection weight of self-organizing radial basis function neural network (SO-RBFNN). Finally, the final output of the model was generated by using the linear combination information from multiple SO-RBFNNs to accelerate the global convergence of the network. To verify the feasibility of the proposed method in practical problems, and test experiments in multiple classification and prediction tasks were conducted, respectively. Results show that this method can not only achieve faster training convergence, but also produce a more concise and compact radial basis function neural network (RBFNN) model. In particular, in the prediction experiment of total phosphorus concentration in wastewater treatment process, the root mean squared error (RMSE) of the test set is reduced by up to 48.90%, achieving better accuracy and stronger generalization ability.