Adaption of counter matrix

After each evaluation cycle the tolerance test vector either seems to be plausible or not. Depending on the result of the test the counter values belonging to each tolerance value are adjusted.

Unconfirmed tolerance test vectors

If a tolerance test vector is not confirmed in any of the dimensions, the current tolerance hypothesis should not enter future tolerance test vectors. Its probability is decreased by reducing the respective counter values. The amount of reduction is prescribed by the adaption rule stated prior to entering the procedure, e.g. reducing the counter value by one fourth. Always, a minimum value of 1 remains. Unlike confirmation of test vectors, not only counter values in specific dimensions are modified. If a tolerance test vector is not confirmed in any dimension at all, the counter values belonging to all current tolerance hypothesis are reduced.

Based on the FACTORFINDER relevance model, how can be explained if there are no plausible reference examples available. It is one of the basic assumptions that the candidate example uses the same ideal as at least one of the reference example. If this is the case, the tolerance of non-ideal values is lower than the test vector suggests. The factors are more important. Or - this possibility must be considered as well - the candidate belongs to another ideal than all other successful examples.

In both cases it is useful to reduce the respective counter values of the test vector. In the first case, because the tolerance was overestimated and in the second case, because this would indicate an extremely different example and therefore warns the analyst. The possibility to pick an example from an unknown ideal was excluded by assumption as the example base is assumed to be complete. Following this assumption, the candidate must use the same ideal as at least one other example. Because of that, the sum of the tolerance hypothesis of the differing dimensions has to be with at least one reference example below 1.0. As this is not the case, the current test vectore must be false.

Confirmed tolerance test vectors

If at least one reference example confirms the test vector in at least one dimension, the list of confirmed tolerance hypothesis is the starting point for the adaption of the counter matrix. One by one the dimensions are processed: If a dimension shows the same value as the candidate and all of the successful reference examples, the counter values of this dimension remain unchanged. The next dimension is selected. If at least one reference example differs from the candidate, inferences are permitted and the counter value that belongs to the tolerance test hypothesis is changed. The counter value is increased depending on the adaption rule that is determined prior to entering the procedure.

Sums of the tolerance hypothesis below 1.0 suppport the test vector. It is plausible. The FACTORFINDER-Procedure modifies the counter matrix in those dimensions, where the tolerance hypothesis is supported by any of the pairs of candidate and reference example. The procedure changes the representation of the tolerance in this factor by enlarging certain counter values. These counter values are those that belong to the confirmed tolerance hypothesis. Consequently, these tolerance options will be more likely selected in future tolerance test vectors.

Only those counter values are modified where the candidate example differs from the reference example. It would be as well possible to change the counter values in all dimensions. By that, tolerance options of dimensions that do not differ would get an increased counter value. As these tolerance options have not been tested in the current cycle, prior adaptions would dissipate within senseless value adjustments. That is the reason why those dimensions where candidate and reference example are identical, the counter values do not change.

Informal inferences on relevance

Factors seem to be important, if all successful variants show the same value and all inacceptable examples do not show this specific value.

Transmissions in stereo are very important for a radio station. This is the impression, if all successful radio stations broadcast in stereo and no unsuccessful station transmits in stereo. If stereo transmissions are found in successful stations and in unsuccessful, this factor seems to be less important.

A factor looks unimportant, if the same value is found with acceptable examples and with inacceptable examples. Although these inferences seem obvious, the procedure can not draw any further benefit from these informal considerations.