Release Date: 8-May-1995
Affected Versions: Neuralyst 1.40b or earlier
Symptoms: Creating a neural network with a total number of target or output cells greater than 32,767 and training it causes an error message SQRT - domain error, subsequently followed by miscalculation of the RMS Error and the number of data items right or wrong. For example, defining a 14 column, 3000 row target area with a corresponding output area, results in 42,000 target or output cells and training this causes the previously described errors.
Problem: Neuralyst uses a 16 bit integer to compute the total number of target or output cells which is then used to calculate RMS Error, data item count, etc. Counts greater than 32767 overflow the 16 bit integer and result in negative counts which then results in the square root and other errors. Solution: Only the summary error report resulting from training or prediction is affected by this problem, so in situations where more than 32,767 target or output cells must be defined, then the trained and predicted outputs will still be correct if the error report is ignored. Where possible, avoid defining more than 32,767 cells as calculated by multiplying the target column count by the row count and rows per pattern count. Neuralyst version 1.40c or greater corrects this problem.