skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Multi-Layer Perceptrons and Support Vector Machines for Detection Problems with Low False Alarm Requirements: an Eight-Month Progress Report

Abstract

In this project, the basic problem is to automatically separate test samples into one of two categories: clean or corrupt. This type of classification problem is known as a two-class classification problem or detection problem. In what follows, we refer to clean examples as negative examples and corrupt examples as positive examples. In a detection problem, a classifier decision on any one sample can be grouped into one of four decision categories: true negative, true positive, false negative and false positive. These four categories are illustrated by Table 1. True negatives and true positives are cases where the classifier has made the correct decision. False positives are cases where the classifier decides positive when the true nature of the sample was negative, and false negatives are cases where the classifier decides negative when the sample was actually positive. To evaluate the performance of a classifier, we run the classifier on all the samples of a data set and then count all the instances of true negatives, true positives, false negatives, and false positives. All of the performance metrics in this report are then formed from a combination of these four basic decision categories.

Authors:
; ; ; ; ; ; ; ;
Publication Date:
Research Org.:
Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
Sponsoring Org.:
USDOE
OSTI Identifier:
922310
Report Number(s):
UCRL-TR-227939
TRN: US200806%%223
DOE Contract Number:
W-7405-ENG-48
Resource Type:
Technical Report
Country of Publication:
United States
Language:
English
Subject:
99 GENERAL AND MISCELLANEOUS//MATHEMATICS, COMPUTING, AND INFORMATION SCIENCE; 42 ENGINEERING; CLASSIFICATION; DETECTION; METRICS; PERFORMANCE; PROGRESS REPORT; VECTORS

Citation Formats

Chen, B, Hickling, T, Krnjajic, M, Hanley, W, Clark, G, Nitao, J, Knapp, D, Hiller, L, and Mugge, M. Multi-Layer Perceptrons and Support Vector Machines for Detection Problems with Low False Alarm Requirements: an Eight-Month Progress Report. United States: N. p., 2007. Web. doi:10.2172/922310.
Chen, B, Hickling, T, Krnjajic, M, Hanley, W, Clark, G, Nitao, J, Knapp, D, Hiller, L, & Mugge, M. Multi-Layer Perceptrons and Support Vector Machines for Detection Problems with Low False Alarm Requirements: an Eight-Month Progress Report. United States. doi:10.2172/922310.
Chen, B, Hickling, T, Krnjajic, M, Hanley, W, Clark, G, Nitao, J, Knapp, D, Hiller, L, and Mugge, M. Tue . "Multi-Layer Perceptrons and Support Vector Machines for Detection Problems with Low False Alarm Requirements: an Eight-Month Progress Report". United States. doi:10.2172/922310. https://www.osti.gov/servlets/purl/922310.
@article{osti_922310,
title = {Multi-Layer Perceptrons and Support Vector Machines for Detection Problems with Low False Alarm Requirements: an Eight-Month Progress Report},
author = {Chen, B and Hickling, T and Krnjajic, M and Hanley, W and Clark, G and Nitao, J and Knapp, D and Hiller, L and Mugge, M},
abstractNote = {In this project, the basic problem is to automatically separate test samples into one of two categories: clean or corrupt. This type of classification problem is known as a two-class classification problem or detection problem. In what follows, we refer to clean examples as negative examples and corrupt examples as positive examples. In a detection problem, a classifier decision on any one sample can be grouped into one of four decision categories: true negative, true positive, false negative and false positive. These four categories are illustrated by Table 1. True negatives and true positives are cases where the classifier has made the correct decision. False positives are cases where the classifier decides positive when the true nature of the sample was negative, and false negatives are cases where the classifier decides negative when the sample was actually positive. To evaluate the performance of a classifier, we run the classifier on all the samples of a data set and then count all the instances of true negatives, true positives, false negatives, and false positives. All of the performance metrics in this report are then formed from a combination of these four basic decision categories.},
doi = {10.2172/922310},
journal = {},
number = ,
volume = ,
place = {United States},
year = {Tue Jan 09 00:00:00 EST 2007},
month = {Tue Jan 09 00:00:00 EST 2007}
}

Technical Report:

Save / Share:
  • This paper presents the Cost-Sensitive Random Subspace Support Vector Classifier (CS-RS-SVC), a new learning algorithm that combines random subspace sampling and bagging with Cost-Sensitive Support Vector Classifiers to more effectively address detection applications burdened by unequal misclassification requirements. When compared to its conventional, non-cost-sensitive counterpart on a two-class signal detection application, random subspace sampling is shown to very effectively leverage the additional flexibility offered by the Cost-Sensitive Support Vector Classifier, yielding a more than four-fold increase in the detection rate at a false alarm rate (FAR) of zero. Moreover, the CS-RS-SVC is shown to be fairly robust to constraints onmore » the feature subspace dimensionality, enabling reductions in computation time of up to 82% with minimal performance degradation.« less
  • The rate at which a mine detection system falsely identifies man-made or natural clutter objects as mines is referred to as the system's false alarm rate (FAR). Generally expressed as a rate per unit area or time, the FAR is one of the primary metrics used to gauge system performance. In this report, an overview is given of statistical methods appropriate for the analysis of data relating to FAR. Techniques are presented for determining a suitable size for the clutter collection area, for summarizing the performance of a single sensor, and for comparing different sensors. For readers requiring more thoroughmore » coverage of the topics discussed, references to the statistical literature are provided. A companion report addresses statistical issues related to the estimation of mine detection probabilities.« less
  • Median filtering reduces speckle in synthetic aperture radar (SAR) imagery while preserving edges, at the expense of coarsening the resolution, by replacing the center pixel of a sliding window by the median value. For shadow detection, this approach helps distinguish shadows from clutter more easily, while preserving shadow shape delineations. However, the nonlinear operation alters the shadow and clutter distributions and statistics, which must be taken into consideration when computing probability of detection and false alarm metrics. Depending on system parameters, median filtering can improve probability of detection and false alarm by orders of magnitude. Herein, we examine shadow probabilitymore » of detection and false alarm in a homogeneous, ideal clutter background after median filter post-processing. Some comments on multi-look processing effects with and without median filtering are also made.« less
  • A detailed parametric study of the major variables that affect the pressure profile in the geothermal well, and hence its performance, was conducted in order to identify and define the pertinent problems that require further experimental and theoretical investigation. The well and, in particular, its two-phase section were modeled on the basis of sound two-phase flow principles. A computer algorithm was developed for pressure profile calculations and simulation of well performance. The algorithm was used to study the effects of total mass flow rate, heat losses to the surrounding formation and void fraction on the pressure profile and the vapormore » and liquid production rates. In addition, several existing void fraction and frictional pressure drop correlations were evaluated in order to determine their applicability to geothermal systems. (MHR)« less
  • The effects of radiation damage, as measured by mutation rates at specific visible autosomal loci, upon the male germ cells of Drosophila was studied with special emphasis upon comparative mutation rates of gametes in different stages of maturation at the time of radiation. Information on the brood patterns. and data on the kind and incidence of mutations as related to the brood patterns are summarized. A clearly defined reduction in number of offspring was found to occur in the third brood (6 to 8 days following radiation) although the number of offspring is also considerably reduced in the fourth brood.more » The data point to a difference in mutation patterns as well as mutation rates for the three types of mutations. The numbers are large enough for the first two broods to make it highly probable that the difference between the mutants at specific loci, and the phenotypic dominaitts is more than just a quantitative difference. In respect to the gonadal periods, it appears that the mutation rate of the Minutes may be considerably higher than of the other types of mutations. lt is also suggested that there may be a difference in mutation rates in gonadal cells in different developmental stages at the time of radiation. (P.C.H.)« less