Abstract
In human learning, we firstly memorize the given data as it is, but we gradually come to extract and store abstract ones. We have simulated this learning, where all data were stored to generate a rule set and evaluate knowledge representations, knowledge sets and reasoning methods. However it is not realistic for human learning. We, therefore, proposed some criteria that the system removes the most unnecessary data when the data exceeds the specified size, but did not study enough. The proposed include the time, correctness and similarity criteria. In this paper, we apply the criteria to the different kinds of data. Then we propose a new criteria in which three criteria are combined. We simulate for the different three kinds of data such as random data, data periodically shifted classes and data with unbalanced number of classes and the new criteria keeps rather high correct rate in all kinds of data.
