Performance evaluation of fusing two different knowledge sources in ripple down rules method

Knowledge acquisition is generally meant to be an action of eliciting knowledge from human experts. On the other hand, knowledge acquisition from data is called machine learning. These two are studied by separate research communities. We have proposed a method to utilize these two different knowledge sources and fuse them into an operational classifier under a framework of Ripple Down Rules (RDR) method. The method is further extended to a situation where an environment changes over time. The principle that unifies all of these is minimum description length principle. In this paper we report the performance evaluation of our method for two kinds of situations where: 1) the knowledge source is changed from the expert to data and vice versa at any time, and 2) both the knowledge source and environment is changed. Experiments were conducted to simulate building RDR trees for the above two situations using the datasets in UCI repository (with appropriate modification to simulate the environment change). The results are encouraging and indicate that our method works well in a situation in which the changes of the knowledge source and environment are coupled.