Michael Bakich

Biography:

Michael Bakich, Ph.D., Electrical Engineering, Purdue and Texas A&M

Abstract:

We first present an error surface in R^3in order to visualize the effect on robustness of algorithm design and uncertainty in input operating conditions and to understand how the slope of such a surface can impact the measure of trust placed in results as a form of operational “go/no go” gauge when in actual battlespace conditions. We then examine the general case in R^n where n>3 where visualization in the standard sense is not possible.

Visualization of global measures of robustness with development in R^3.

Without considering “goodness” of algorithm design there will obviously be variability of ATR performance as a function of Operating Condition variability itself. This is understood as uncertainty and is represented by a Gaussian distribution of probabilities rather than as a fixed certain value. As the actual value on input deviates from an assumed value, there is a propagation of error. When considering the global output of the algorithm, this results in an error surface. It is instructive to understand robustness as a surface representing Mean Square Error. We characterize this as uncertainty and represent this as a Gaussian distribution about a proposed or nominal value. By visualization of MSE as a surface we may easily demonstrate that even under the most simple of shifts in operating conditions, there can be performance altering changes from the world of the algorithm design and the world of the real battlefield. We note that the output of an algorithm designed to determine a result from sensor data propagates this uncertainty and it is manifest in the sensor/algorithm output. We employ the concept of the Gaussian correlation matrix then to calculate MSE between an actual output (truth) and this result. This error is inversely proportional to a level of “trust” in output results for this system.

Understanding global measures of robustness with development in R^n.

To avoid difficulties visualizing levels of error in higher dimensions we will employ a histogram method to convey essential results. Without considering “goodness” of algorithm design there will obviously be variability of ATR performance as a function of OC variability itself. We characterize this as uncertainty and represent this in a simple example as a Gaussian distribution about a proposed or nominal value. In this paper we demonstrate that even under the most simple of shifts in operating conditions, there can be performance altering changes from the world of the algorithm design and the world of the real battlefield. We note that the output of an algorithm designed to determine a result from sensor data propagates this uncertainty and it is manifest in the sensor/algorithm output. We employ the concept of the Gaussian correlation matrix then to calculate MSE between an actual output (truth) and this result. This error is inversely proportional to a level of “trust” in output results for this system.