Human activity recognition (HAR) systems combined with machine learning normally serve users based on a fixed sensor position interface. Variations in the installing position will alter the performance of the recognition and will require a new training dataset. Therefore, we need to understand the role of sensor position in HAR system design to optimize its effect. In this paper, we designed an optimization scheme with virtual sensor data for the HAR system. The system is able to generate the optimal sensor position from all possible locations under a given sensor number. Utilizing virtual sensor data, the training dataset can be accessed at low cost. The system can help the decision-making process of sensor position selection with great accuracy using feedback, as well as output the classifier at a lower cost than a conventional training model.