A machine learning method designing flat broadband erbium-doped fiber amplifier (EDFA) + Raman hybrid amplifier was demonstrated. First, we trained a neural network (NN) using data consisting of Raman amplifier pump parameters and gains arranged in descending, ascending, and random order. This trained NN is utilized to calculate pump parameters based on the target gain of the hybrid amplifier. Then, the gain accuracy is optimized through the application of fitting equations that establish a relationship between the adjustment power of Raman’s pump and the gain ratio of the EDFA/Raman amplifier. The optimized results show that when using descending datasets, the error of output power is minimized, with means of root-mean-square-error and maximum error of 0.24 and 0.54 mW, respectively. Its average flatness is not significantly different from the results obtained using ascending and randomly arranged training data. Second, we divided the training data into five parts according to gain ratio of EDFA and Raman amplifier. The accuracy deteriorates when training the NN with a portion of the data, but the accuracy could be enhanced by repeatedly training the NN with the same data. Our results contribute to the construction of intelligent fiber optic transmission networks.