One of the models of artificial neural network supervised learning favorite is a model of backpropagation learning. The performance of the algorithm is influenced by some network parameters including many neurons in input layer, maximum epoch used, the magnitude of the learning rate, and the resulting error (MSE). The performance of the training algorithm is said to be optimally viewed from the resulting error.The less resulting error, the more optimal the performance of the algorithm. Tests conducted in the previous research found that the most optimal training algorithm based on smallest error result is the Levenberg - Marquardt training algorithm with an average of MSE 0.001 with the level of testing α = 5%. In this research was conducted to determine the most neuron in hidden layer viewed from epoch necessary to reach smallest error condition. This research uses mixed method which is development research with quantitative and qualitative test (using ANOVA statistic test). The research data were taken from random data with 5, 10, 15 input neurons, n neurons in the hidden layer {n = (1, 2, 3, 4, 5, 7, 9)(6, 7, 8, 9, 10, 12, 14, 16, 19)(8, 11, 13, 15, 17, 19, 21, 23, 27, 29)} and 1 neuron in the output layer. The results of the analysis show that 5 input neuron is 9 neuron , with epoch average 10.80 in learning rate = 0.01. In the 10 input neuron is 19 neuron, with epoch average 21.52 in learning rate = 0.7. Meanwhile in 15 input neuron is 29 neuron, with epoch average 7.38 in learning rate = 0.7.