Local Position Estimation Using an Artificial Neural Network Based Model with a Hardware Implementation

Efficient implementation of the activation function is an important part in the hardware design of artificial neural network. Sigmoid function is one of the most widely used activation function. In this paper, an efficient architecture for digital hardware implementation of sigmoid function is presented. The proposed method used second order nonlinear function (SONF) as a foundation and further improves the result by using 320 bits of read only memory (ROM) for storing a differential lookup table (differential LUT). The method proves to be more effective considering the smallest deviation of sigmoid function achieved in comparison to conventional LUT and SONF. Employing this method for hardware-based ANN in the indoor positioning system have shown that, ANN can detect the target position almost as accurate as software implementation with a speed 13 times faster. Thus the proposed idea is suitable to be implemented in a hardware-based ANN for various real-time applications.