Positive and negative patterning in human causal learning

Abstract

Investigations of patterning discriminations by nonhuman animals have generally found that positive patterning is easier to learn than negative patterning. Studies of patterning discriminations in human causal learning tasks have failed to document any differences between positive and negative patterning. In the present study, human participants predicted an outcome on trials involving either a compound cue or its elements. Positive and negative patterning problems were successfully solved in a within-subjects design; negative patterning problems proved to be more difficult when an additional, 50% contingent cue was included (Experiment 2), but not when it was excluded (Experiment 1). Possible reasons for these results are discussed. The discussion concludes with an analysis of exemplar models (e.g., Pearce, 1994) of human causal learning and considers the conditions under which these models do and do not anticipate our results.

Asymmetry of generalization decrement in causal learning

“…In a negative patterning procedure A+ and B+ trials occur alongside unreinforced trials with the compound (AB-trials). The discrimination between A, B, and the AB compound can be reliably acquired in animals and in humans (Rescorla, 1972;Shanks, Charles, Darby, & Azmi, 1998;Young, Wasserman, & Johnson, 2000). However, this result cannot be accommodated in an elemental model that represents the AB compound simply as the joint presence of A and B; if that were the case then AB should provoke stronger responding than that to either A or B alone through a process of summation.…”

When One Cue is not Enough: Combining Fast and Frugal Heuristics with Compound Cue Processing

“…In the XOR (eXclusive OR) problem, the probability of the effect was greater when one cause alone was present (100%) than when the two causes appeared (0%) or none appeared (0%). In both problems, participants were able to learn to predict the effect based on the type of cue interaction presented (Young et al, 2000; see also Edgell, 1993;Lachnit & Kimmel, 1993;Shanks, Charles, Darby, & Azmi, 1998a;Williams & Braker, 1999).…”

“…In contrast, the AND problem, which is partially configural, 1 has a linearly separable structure (Zurada, 1992). Linearly separable problems have been shown to be easier to learn than nonlinearly separable problems in some experiments (Kimmel & Lachnit, 1991;Lachnit & Kimmel, 1993;Smith, Murray, & Minda, 1997) but not in others (Shanks et al, 1998a;Shanks & Darby, 1998;Young et al, 2000).…”

“…Most of the experiments that analysed compound cue processing did so within the range of two or three cues that are combined in one compound (e.g., Edgell, 1993;Young et al, 2000). In real environments, however, there exists a multitude of potential combinations of cues to form compounds, which makes it virtually impossible for humans to keep track of all of them.…”

“…For example, in a typical causality learning study by Young, Wasserman, Johnson, and Jones (2000), participants were given two candidate causes (i.e., two substances, appearing either alone or in conjunction with one another) for an effect (i.e., the presence of a disease in a sample of patients). The participants' task was to infer the causal relationship between the candidate causes and the effect.…”