Sonar sensitivity analysis using a neural network acoustic model emulator

A technique is reviewed for training artificial neural networks to emulate the complicated input-output relationships of an acoustic model. This neural network acoustic model emulator is intended for use in a sonar controller, which may require a large number of forward model runs to determine the optimal sonar setting in a given environment. The neural network can supply sonar performance predictions to high enough fidelity for use in a controller, but with a much reduced computational burden compared to the original acoustic model. Among the challenges of developing control guidelines for highly variable littoral areas is the difficulty in understanding the sensitivity of acoustic response to small changes in environmental or sonar control parameters. An effective sensitivity analysis tool would allow users or automatic control algorithms to place a control emphasis on those parameters that have the greatest effect on sonar response. Additionally, an improved understanding of acoustic sensitivity may lead to improvements in model and controller development. In this paper, the neural networks, originally developed for automatic sonar controllers, are used to explore the sensitivity of the system. Given a properly trained neural network, sensitivity measures can be directly calculated. The neural networks can also be used to visualize the effect of changing environmental and control parameters. A variety of ways in which the neural network structures can be used to examine the sensitivity of the sonar system is presented.