probabilities using index-count filtering are larger than those for exponential filtering.
For Titan, the results are mixed, further suggesting that Titan reliability has not
improved in recent years.

For comparison purposes, the same filtering techniques have been applied to all flight
tests shown in the tables of Appendix D, regardless of configuration. The results are
presented in Table 3.

              Table 3. Predicted Failure Probabilities for All Configurations
                                          Filter Technic ue                    Sample
              Flight      Equal     Index      Expon.      Expon     Expon Failures
   Vehicle Phase         Weight     Count F =0.99 F=0.98 F =0.97               /Total
   Atlas      0         0         0           0           0         0         0/7
              0-1       0.1053    0.0641      0.0422      0.0273    0.0190    56/532
              0-2       0.1711    0.0990      0.0555      0.0311    0.0204    91/532
              0-3       0.2086    0.1261      0.0802      0.0559    0.0455    111/532
              0-4       0.2143    0.1330      0.0873      0.0627    0.0511    114/532
              0-5 •     0.2575    0.1671      0.1150      0.0866    0.0725    137/532
   Delta      0         0         0           0           0         0         0/196
              0-1.      0.0172    0.0164      0.0148      0.0110    0.0077    4/232
              0-2       0.0259    0.0232      0.0201      0.0133    0.0085    6/232
              0-3       0.0431    0.0279      0.0263      0.0150    0.0089    10/232
              0-4       0.0431    0.0279      0.0263      0.0150    0.0089    10/232
              0-5* 0.1078         0.0766      0.0740     0.0536     0.0459    25/232
   Titan      0         0.0306    0.0137      0.0187     0.0281     0.0349    3/98
              0-1       0.0534    0.0319      0.0351     0.0399     0.0467    18/337
              0-2       0.1424    0.0771      0.0719     0.0662     0.0750    48/337
              0-3       0.1632    0.0924      0.0830     0.0711     0.0770    55/337
              0-4       0.1662    0.0942      0.0840     0.0712     0.0771    56/337
              0-5·      0.1958·   0.1369      0.1326     0.1277     0.1346    66/337
   • Includes response mode 'NA'

.A comparison of Table 2 and Table 3 shows that in most cases, but not all, exponential
 filtering produces failure probabilities for the representative configuration samples that
 are smaller than the corresponding probabilities for the all-configuration samples. The
 fact that most differences between corresponding samples are relatively small attests to
 the effectiveness of the exponential filter in down-weighting early launch failures. This
 is not the case for equal weighting of tests, where the predicted failure probabilities
 based on all configurations are up to 3.6 times as large.

With respect to- the weighting of missile and space-vehicle performance data, RTI
favors an exponential filter over either the equal-weight or index-count filters.
Weighting percentages for the three filters are given in Table 4 for sample sizes of 4 to
1,000. Except for small samples, the percentages produced by equal weighting place
too much emphasis on old data, thus failing to account for the learning process and

9/10/96                                     18                                         RTI