Automated Factor-Based Sensitivity Analysis

WarpIV Technologies, Inc. proposes to develop a Response Surface Methodology (RSM) tool that will intelligently guide excursions of the AWS MD Engage-ability prelaunch FOM algorithm (and any algorithms under test) to reconstruct its response surface using tenth-order polynomials of up to 100 control variables. Brute force polynomial curve fitting techniques cannot be used to solve this problem because the polynomial in its full form would have 4.69 times ten to the 13th power number of terms. Therefore, the focus of this effort will be to carefully guide excursions of the algorithm under test in an iterative manner so that non-contributing terms in the polynomial can be rapidly excluded. It is assumed that the response surface can be accurately represented by no more than a few thousand terms. This assumption is reasonable and is really a requirement for the success of any sensitivity analysis technique that covers the full range of the control-parameter space. In a sense, the strategy in this proposal is comparable to finding a few thousand needles in a haystack with 4.69 times ten to the 13th power number of straws of hay. Polynomial representations are ideal for performing sensitivity analysis on an algorithm under test, optimizing its performance, and/or finding Nash equilibriums using game theory to determine enemy intent because: (1) polynomials can be quickly evaluated, and (2) first and second derivatives (used by a variety of optimization algorithms) are easily obtained without requiring numerical approximations. The proposal describes the full algorithm that will be developed in this effort. Portions of it have already been developed by WarpIV Technologies, Inc. in 2019 IR&D efforts. One of the critical features of the proposed RSM tool is its use of the WarpIV Optimization Engine that will manage the execution of excursions in all types of parallel and distributed multicore-computing environments. Highly efficient and robust parallel processing techniques will be used at every stage of processing to eliminate bottlenecks and to take full advantage of all compute resources that might be available to the user. This includes everything from multicore laptops, desktops, networks of machines, clouds, server clusters and the world’s largest supercomputers. Approved for Public Release | 20-MDA-10643 (3 Dec 20)