Prediction and diagnosis: Task variable revealed

The authors' work on demonstrating fundamental differences between predictive learning and diagnostic learning lead incidentally to the discovery of a critical task variable operating in diagnostic tasks. In our experiments these tasks consist of random sampling with replacement of cues from two cue-generators. The subject has to infer from which cue-generator the cue was drawn. Two different sampling procedures can be distinguished. Under forced sampling the predetermined probabilities of the occurrence of each cue agree exactly with the relative frequencies within each block of trials. Under free sampling there are no with in-block restrictions on the sampling, which results in sampling fluctuations from block to block. It was demonstrated that this difference in sampling is a critical task variable. Free sampling results in a higher frequency of long sequences of confirming trials, that is, trials in which feedback indicates the optimal rule. The excess of long sequences of confirming trials under free sampling makes this kind of sampling a better condition for efficient learning than forced sampling. If one believes (as the authors do) that "nature" does not apply forced sampling the results indicate that the "natural" way of sampling is also the most favorable for learning diagnostic tasks.     

Prediction and diagnosis: An experimental comparison

Diagnostic tasks differ from predictive tasks in two respects (1) the direction of the inference in diagnosis (from effect to cause) is inverse to that in prediction (from cause to effect), and (2) the event to be inferred is more distal in diagnostic tasks than in predictive tasks. These task features led to the hypothesis that a diagnostic task would be harder to learn than a predictive task. This hypothesis was tested in an experimental situation in which the causal mechanism was random sampling with replacement. Five groups of subjects with 10 subjects in each (two predictive and three diagnostic groups) were run in individual sessions with 100 trials of feedback training. The two predictive groups learned more rapidly and to a higher level compared to the diagnostic groups. Conclusion: the present data support the hypothesis that predictive tasks are easier to learn than comparable diagnostic tasks.