Higher-order yield functions, such as Yld2000-2D, improve the accuracy of sheet metal forming simulations. Calibrating their parameters requires material testing under biaxial stress states, and biaxial tensile tests using cruciform specimens have been conducted. However, biaxial tensile test machines are not widely used, and the material testing is difficult. Therefore, various specially shaped tensile test specimens have been proposed. An hourglass-shaped specimen was also developed that can generate a wide range of biaxial stress states using a uniaxial tensile testing machine. Uniaxial tensile tests were conducted on mild steel sheets with this specimen, and strain distributions were measured using the digital image correlation method. Machine learning techniques based on multilayer neural networks were used to identify the parameters of the Yld2000-2d yield function. Virtual material parameters were generated, and tensile test analysis were performed. The strain distribution from the analysis and these material parameters were used as training data. The parameters of the Yld2000-2d yield function were estimated from the measured strain distribution of the hourglass-shaped specimen using machine learning. On the other hand, the parameters of the Yld2000-2d yield function were identified from conventional uniaxial and biaxial tensile test results. The yield curves obtained by both methods showed good agreement.