Neural network ensonification emulation: Training and application.

摘要

This dissertation investigates several modifications and extensions of conventional neural networks for application to the problem of optimally choosing the adjustable parameters in a sonar system. In general, neural networks offer several key advantages over other technologies that might be used for this task, including the ability to learn from examples and the ability to extract information about the underlying system through neural network inversion. One aspect of this work is the use of a neural network for emulating a computationally intensive acoustic model. A novel neural network training technique for varying output node dimension is developed, allowing a single neural network to be used for different output topologies. Step size modification for this training technique is also introduced to improve accuracy, convergence time, and the smoothness of the weight space, eventually providing better generalization. Inversion of neural networks is also investigated in order to solve for the optimal control parameters given a requested level of sonar performance. In order to improve inversion accuracy, modular neural networks are designed using adaptive resonance theory for pre-clustering. In addition, sensitivity of the feed forward layered perceptron neural network is derived in this work. Sensitivity information (i.e., how small changes in input layer neurons affect output layer neurons) can be very useful in both the inversion process and system performance analysis. Finally, the multiple sonar ping optimization problem is addressed using an evolutionary computation algorithm applied to the results of properly trained neural networks. It searches for the combination of control parameters over multiple independent sonar pings that maximizes the combined sonar coverage.