Abstract:

We propose a novel algorithm for extracting rules from artificial neural network. After the network is trained and pruned successfully, the activation values at the hidden unit are clustered into discrete values. In the cluster phase, the cluster number of weights can be adjusted dynamically according to activation values of their corresponding hidden units. The experiment results have shown its feasibility and accuracy, and prove that this algorithm decreases the complexity of rules extraction and makes the number of the extracted rules small.

Key words: extracting rules, artificial neural network, clustering