This paper proposes an interactive training system for control of myoelectric prostheses. The proposed training system is capable of selecting suitable motions (EMG patterns) for each user by eliminating ineffective ones, and can also provide consistency between user's motor images and the corresponding prosthetic movements using a virtual prosthetic hand (VH). In the experiments performed, a one-day training session using the proposed system was conducted with nine healthy males (including an experienced) and an upper limb amputee. In addition, EMG discrimination ability of each subject during VH control without any feedback information was evaluated before and after the training to verify the training effects of the proposed system. The results showed that the discrimination rates for selected motions were sufficiently high (98.9 ± 1.24%) by using the proposed selection method, and the accuracy in discrimination for VH control was significantly improved after training (for healthy subjects and the amputee at the 0.1% and 1% level, respectively). It is therefore confirmed that the proposed system can be used for myoelectric prosthesis control training.