Various attempts towards training nets to represent rule based systems or extracting rules from neural nets have been tried. Most of the previous work was concentrating on standard rule structure of the form IF <condition> THEN <action>. In this paper we present two methods to train a neural net to correctly represent censored production rules of the form IF <condition> THEN <action> UNLESS <censor>. The advantage of doing this over the standard rule structure is that censor production rules can very well serve in real time systems. Using censor production rules allow us to get more certain results given more time by checking more censors which rarely occur. One of the methods is based totally on backpropagation with a small modification. The second method is partially based on backpropagation and the rest is based on a proposed algorithm that is concerned with adjusting the net weights taking into account the importance of the censors. The weights of the links connecting the censors with the hidden layers represent the time allocated to each censor to be checked in time constraints. These weights are based on the importance of the censor and the average time for censor checking. We also present a method to extract the rules from the trained net.

Keywords:

Rule-based systems, Variable precision systems, Neural networks