Ballistic gun fire control using a feedforward network with hybrid neurons

In this paper, we discussed training a network with hybrid neurons consisting of both sigmoidal and linear neurons for the application of real-time ballistic gun fire control. Also, we developed a training algorithm for such networks, modifying the traditional gradient descent algorithms. Unlike the traditional ones, the error, fed back in order to adjust the connection weights, is controlled in a manner to attenuate too big errors and maintain small errors. This error control turns out to make a significant improvement in training networks with linear output neurons. The networks with linear neurons in the output layer have many advantages, such as efficient training, catering to linear functions, obviating the need for the external de-scaling of the final output after training, and so forth.