Adaptive Gridding in Complex Physical Environments to Reduce Uncertainty

Effective Naval warfare depends on accurate estimates of sensor performance in complex environments. Models for sonar performance provide a force multiplier by allowing own ship to optimize its tactics while avoiding counterdetection. The shift in emphasis from open ocean to littoral operations has increased the demands on models of sonar performance because data must be sampled at an order-of-magnitude finer resolution to obtain the same accuracy in littoral regions as in the open ocean. This requirement makes many models impractical to use in time-sensitive operations. A traditional response to this problem has been to replace detailed physical simulations with simpler, sensor specific, heuristics. This approach is inflexible and generally provides inferior predictions. The solution to the problem of obtaining accurate sonar performance estimates in time to meet operational constraints is to run detailed physical models at a carefully selected set of grid points. This project will use Bayesian Neural Networks to select a minimal number of grid points where detailed acoustic models will be run and which will allow the Bayesian Neural Network to make accurate predictions of sonar effectiveness throughout an operational area. The predictions will be accompanied by estimates of their accuracy.