During the first experiment we demonstrated a direct effect of similarity: subjects are more likely to select a toy to give to the Fennimal if this toy previously resulted in positive feedback when given to a similar Fennimal. In the second experiment we demonstrated an indirect effect of similarity: subjects are more likely to give a toy to the Fennimal if this toy previously resulted in positive feedback when given to a Fennimal that shares overlap to a Fennimal sharing overlap to the observed Fennimal.
The third experiment is conducted for two reasons. First, we want to rule out an alternative explanation to the results obtained in experiment 2 (demonstrating the indirect effect). This alternative explanation holds that subjects learned to associate which sets of two toys "go together" based on the overlap in Fennimals during the training phase of the experiment. If this explanation held true, then during the test phase of the experiment subjects do not conduct a search of their memory, but instead randomly select one of the two clusters of toys to give to the Fennimal. For one cluster both actions were available (and the subject would randomly pick between them if this cluster is selected). For the other cluster, only one toy is available (the action predicted by the CASES model). Under this strategy the subject would be 50% likely to give the CASES-predicted toy to the Fennimal (which is what was observed during experiment 2). Second, the third experiment is aimed at a self-replication of the effects obtained from the first two experiments.
In experiment 3 we use an identical experimental design as in the first two experiments, with the exception of different training and test sets of Fennimals. In particular, during this experiment we will modify the training and test-set Fennimals such that three Fennimals contain partially overlapping features. The training set will consist of 4 Fennimals. Here, 3 Fennimals will share partially overlapping features: Fennimals 1 and 2 will share a type (but have different colors) and Fennimals 2 and 3 will share a color (but be of different types). Fennimals 1 and 3 do not share any overlap in features. Fennimal 4 will not share any overlap to any of the other Fennimals. All Fennimals will have a positive association to a different toy.
The test-set of experiment 3 will consist of 4 trials. Two of these will be “direct trials”, in which the Fennimal will either share a color or a type with Fennimal 4 from the training set. In these trials all toys are available to the subject, and the CASES model predicts the subject to select the toy previously associated to Fennimal 4. The other two trials will be “indirect trials”. In one of these trials the subject will see a Fennimal that shares a color with Fennimal 1 (and has a unique type). Here the toy associated with Fennimal 1 is not available, and the CASES model predicts the subject to select the toy previously associated with Fennimal 2. In the other indirect trial the subject will see a Fennimal that shares a type with Fennimal 3 (and has a unique color). Here the toy associated with Fennimal 3 is not available and the CASES model predicts that the subject selects the toy previously associated to Fennimal 2.
This experimental design allows us exclude the possibility of a statistical-clustering strategy. As Fennimals 1, 2 and 3 now belong to the same cluster, and the toys associated to two of these Fennimals are available to choose during the indirect trials (only one of these being predicted by the CASES model), then the subject can no longer rely on a statistical-clustering approach to select a single action.