Most of the recent studies concerned with mediation in shift learning have utilized one common core concept, that is, dimension, regardless of the various meanings attached to mediational processes. For Kendlers, mediating events arr assumed to be coordinated to introspective reports, language behavior, muscular movement, and other observable events. Especially they emphasized the importance of verbal processes in human concept learning and then the mediating mechanism has been referred to the representational responses associated with a relevant stimulus dimension. Hence judging from the fact that the locus of mediation is attributed to a stimulus dimension, Kendlers' formulation would be categorized as a dimensional mediation theory. Nevertheless, as some investigators stated before, dimensional theory could not cover the findings that even unrelated stimuli serve to rapid reversals. Therefore this paper is an attempt at introducing the mathematical set theory to discrimination shift learning and thereby providing an al. ternative model in this field to develop the concept of mediation. Two fundamental hypotheses were proposed as follows; 1)Human adults could attend to the stimulus set connected with reward and its complementary set, and then utilize the mutually exclusive rule between these sets through responding to stimuli. 2)Infants and young children may forcus on only the dominant stimulus set associated with reinforcement and pay no attension to the other stimuli. In order to affirm the first of these basic hypotheses within the reversal-nonreversal framework, two working hypotheses were made for the experiment as follows; (1)The reversal shift(RS)task could be attained more easily than the nonreversal shift(NRS)task or the control(C)task, for under the RS condition the same stimulus sets hold throughout learnings. (2)There would be no difference in learning between the NRS task and the C task, for under the either tasks the stimulus sets must be changed independently. Figure-1 illustrated the basic diagrams of this experiment, which consisted of two paired-associate learning sessions analogous to the shift experiment. Thirty college students were employed as Ss. The stimuli were Japanese nonsense bigrams and responses single digit numbers. The standard paired-associate learning procedures were used. After completion of learning, Ss were questioned what sort of learning strategies they took. This information was used to classify the learning types Ss followed(TABLE-6). Results were as follows; 1)Working hypotheses were significantly confirmed(TABLE-4, -5). 2)Classification of learning types Ss took showed that mediational learners(SR type)attained the paired-associete task more easily than other non-mediational learners(ER type and S-R type)(TABLE-7, -8). In conclusion, considering the above findings, this experiment confirmed not only the working hypotheses but also the first of the fundamental hypotheses.