The situation decomposition extracts multiple situations, each of which is a combination of an attribute set and a case set, from relational data. The characteristics of extracted situations depend on the choice of evaluation criterion to situations, because situation decomposition algorithms search for the local maximums of a criterion value. It is already known that situation decomposition using a Matchability criterion has a high prediction performance for card classification tasks using a small number of training cases. However, the performance is still inferior to human's ability. In this paper, an ETMIC (Edge of Total Mutual Information Cliff) criterion is proposed as a new criterion for evaluating situations. The situation decomposition using the ETMIC criterion extracts situations, which have three tendencies: a high covariant relationship inside, few influences from unselected attributes, and a large number of cases. The prediction system using the ETMIC situation decomposition can correctly answer 98% to the card classification task (which needs learning disjunctive concepts) using only 20 training cases. The prediction ability of the proposed system extremely exceeds that of the conventional one. The change of the acquisition rate of the concept to the number of experience cases is evaluated on the card task simulation with the ETMIC situation decomposition. The result is consistent with data from a psychological experiment.