It was seen that for most parameters, there is a lot of difference between the influence had by colour on physical and chemical parameters and had on colour by those same parameters. It was expected that there would be broad agreement between the relative connection weight influence of any one physical or chemical parameter on colour and vice versa, but this is not the case. This disagreement between the two network types is suspected to be due to the fact that the average connection weight influence was taken across several networks in each case, with some parameters
predicted relatively inaccurately. This could result in high levels of variance for parameter relationships amongst those parameters that are less successfully predicted. An example would be extractable manganese, which is not very well predicted by colour (Tables 2 & 3) but which appears to have a strong influence on colour (Figs. 1 and 2). The neural network models predicting colour from chemical composition were unable to produce consistent high or low connection weights from extractable manganese as an input parameter, and the initial randomisation of the models was therefore relatively unchanged. However, when applied to parameters that are predicted with more accuracy, we find that relatively high predictive ability does not imply strong relationships between input and output parameters. A high level of relationship strength variation was seen across the 127 networks used, resulting in mean values that are not meaningful. This indicates that the sensitivity analysis of the neural network models was not as useful in this case as it has been in others, and that other approaches should be used for multiple networks.