last modified: 19-NOV-1985 | catalog | categories | new | search |

NESC1002 SCREEN.

SCREEN, Statistical Sensitivity Ranking of Program Input Variables

top ]
1. NAME OR DESIGNATION OF PROGRAM:  SCREEN.
top ]
2. COMPUTERS
To submit a request, click below on the link of the version you wish to order. Only liaison officers are authorised to submit online requests. Rules for requesters are available here.
Program name Package id Status Status date
SCREEN NESC1002/01 Tested 19-NOV-1985

Machines used:

Package ID Orig. computer Test computer
NESC1002/01 IBM 3033 IBM 3084Q
top ]
3. DESCRIPTION OF PROGRAM OR FUNCTION

SCREEN is a statistical sensitivity analysis procedure for ranking input data of large computer codes in the order of sensitivity importance. The problem is to determine a group of the most important input parameters of a  computer code when the total number of input variables is large, so  large that standard sensitivity evaluations varying each input variable (one or two at a time) would be prohibitively expensive.
top ]
4. METHOD OF SOLUTION

SCREEN selects values for the input parameters  of a deterministic computer program from input-specified regions of  interest. The regions are defined by probability distributions and confidence intervals. The program arranges these input values into combinations called 'knot-points' that serve as input for the deterministic code that calculates the values of interesting output  variables (consequences) for the specified knot-points.
The output/consequence values for the knot-points are then used to determine (1) which input variable are most correlated with the output/consequence values, using stagewise correlation analysis, (2) which input variables are most likely to contribute to discontinuities or threshold effects in the output values, using statistical tests for subset characteristics, and (3) which group of input parameters yields the best significant regression model, using quadratic models and successive regression analyses with an increasing number of parameters. The regression part of SCREEN evaluates all regressions of two variables using the principle of stepwise regression analysis for a large number of variables. Both residual errors and special sensitivity/spuriousness indices can be  used to select seed input variables for each step of the regression  analysis. The significance of each added parameter of a model can be assessed by F-statistics for regression models. Student t-statistics and extreme value statistics are used for testing threshold effects.
top ]
5. RESTRICTIONS ON THE COMPLEXITY OF THE PROBLEM

Maxima of -
   1000 input parameter variables
      6 output/consequence variables
When selecting the cases to be run (knot-points), eight optional distributions are available for the input parameters, including uniform, exponential, normal, truncated normal, log-normal, and beta distributions.
top ]
6. TYPICAL RUNNING TIME

Running time depends strongly on the size of  the problem, i.e., number of input parameters, knot-points, and the  number of regression steps needed to complete the screening. For small problems where the number of input parameters and knot-points  are both less than 50, the total running time is typically a few CPU seconds on an IBM370/195. Large problems, with approximately 500 input parameters and 50 knot-points, take several minutes of CPU time per consequence on an IBM370/195. NESC executed the sample problem in 2 CPU seconds on an IBM3033.
NESC1002/01
NEA-DB executed the test case included in this package  on IBM 3084Q in 2 seconds of CPU time.
top ]
7. UNUSUAL FEATURES OF THE PROGRAM

Compared to other screening techniques available, SCREEN has the following somewhat unique features: (a) the values of the input parameters (knot-point coordinates) are selected from a continuous distribution rather than from discrete levels; (b) quadratic regression models rather than linear models are used; (c) the total number of input variables may  be much larger than the number of cases and no prior elimination of  variables by judgment is necessary; (d) the regression analysis in SCREEN is more extensive than the standard stepwise regression analysis procedure; and (e) extreme value and t-tests are used to identify parameters important to threshold (discontinuity) effects.
top ]
8. RELATED AND AUXILIARY PROGRAMS

The SCREEN code can be used in conjunction with any separate deterministic code (typically an accident-analysis code such as SAS3D) that provides data for screening. It can also be used to generate input data in the correct format for the response-surface analysis code PROSA2 (NESC Abstract  778).
top ]
9. STATUS
Package ID Status date Status
NESC1002/01 19-NOV-1985 Tested at NEADB
top ]
10. REFERENCES

- J.K. Vaurio,
  Response Surface Techniques Developed for Probabilistic Analysis
  of Accident Consequences, Proceedings of the American Nuclear
  Society Topical Meeting on Probabilistic Analysis of Nuclear
  Reactor Safety, Los Angeles, California, May 8-10, 1978.
- J.K. Vaurio,
  Methods for Statistical Determination of Effective Input
  Variables,
  Transactions of the American Nuclear Society, Vol. 32, pp.
  296-297, 1979.
- J.K. Vaurio,
  Statistical Determination of Threshold Variables,
  Transactions of the American Nuclear Society, Vol. 35, pp.
  263-264, 1980.
NESC1002/01, included references:
- J.K. Vaurio:
  Statistical Identification of Effective Input Variables.
  ANL-82-57  (September 1982)
top ]
11. MACHINE REQUIREMENTS:  250K bytes of memory are required for execution.
NESC1002/01
The test case was run on IBM 3084Q in 2ooK bytes of main storage.
top ]
12. PROGRAMMING LANGUAGE(S) USED
Package ID Computer language
NESC1002/01 FORTRAN+ASSEMBLER
top ]
13. OPERATING SYSTEM UNDER WHICH PROGRAM IS EXECUTED:  OS/MVT, MVS.
top ]
14. OTHER PROGRAMMING OR OPERATING INFORMATION OR RESTRICTIONS

SCREEN
contains some FORMAT statements which use the T edit descriptor and  READ statements which use the ERR= and END= error and end-of-file specifiers. Subroutines ABEND, ALLOC2, FLTRNF, FREE2, and LOC are written in Basic Assembly Language. FLTRNF, ALLOC2, and FREE2 are Argonne National Laboratory computing environment routines. The FLTRNF function statement U=FLTRNF(0) returns uniform random numbers U, between 0 and 1. Subroutines ALLOC2 and FREE2 dynamically allocate and release space for the arrays used in the regression analysis. Subroutine SASIN, which prepares input for an external deterministic code that calculates consequences, is to be supplied by the user. Versions of SASIN appropriate for the VENUS2 and SAS3D  computer programs are included in the package.
top ]
15. NAME AND ESTABLISHMENT OF AUTHORS

     J.K. Vaurio
     Reactor Analysis and Safety Division
     Argonne National Laboratory
     9700 South Cass Avenue
     Argonne, Illinois 60439
top ]
16. MATERIAL AVAILABLE
NESC1002/01
File name File description Records
NESC1002_01.003 INFORMATION FILE 53
NESC1002_01.004 SCREEN SOURCE PROGRAM (FORTRAN) 4570
NESC1002_01.005 SCREEN ASSEMBLER ROUTINES 344
NESC1002_01.006 SCREEN JCL TO RUN SAMPLE PROBLEM 35
NESC1002_01.007 SCREEN SAMPLE PROBLEM INPUT DATA 179
NESC1002_01.008 SCREEN SAMPLE PROBLEM PRINTED OUTPUT 2003
NESC1002_01.009 SUBROUTINE SASIN FOR SAS3D PROGRAM 928
top ]
17. CATEGORIES
  • P. General Mathematical and Computing System Routines

Keywords: accidents, correlations, least square fit, probability, randomness, regression analysis, risk assessment, safety, sensitivity analysis, statistics.