The present paper proposes a novel identification method (RL-BP) for miniature unmanned aircrafts, utilizing Reinforcement-Learning Algorithm to explore the unknown environment, thus optimizing to the appropriate hidden layer node number of the neural network. RL-BP then constructs the corresponding network, trains through samples and updates the network weights, wherein the reward function values are fed back to Reinforcement-Learning Algorithm for optimization. This paper represents and analyzes the RL-BP method, and verifies the method with recorded flight data. The test results show that RL-BP greatly improves upon traditional neural network identification method in both resource consumption and computation accuracy, as RL-BP reduces Average Relative Error by 37.89% and Maximum Relative Error by 31.44% on an average.