Outcome Predictability: Does associative history of outcomes bias subsequent learning in a human goal-tracking paradigm?

In 1975, Mackintosh proposed that a cue previously experienced to be a better predictor of the outcome than the other cues present possesses greater associability. More recently, a study using a human causal learning task demonstrated better learning about the outcome which has been consistently predictable in the past, as compared to the outcome previously experienced to be unpredictable, namely the outcome predictability effect (Griffiths, Mitchell, Bethmont and Lovibond, 2015). The present study aimed to examine the generality of the effect with a novel goal tracking paradigm and determine if the learned predictability of an outcome can shape the associability of this outcome to be entered into novel associations, which is similar to the Mackintosh theory proposed for cues. Seven experiments were conducted by approaching three different designs to manipulate outcome predictability. For the first four experiments, one outcome was consistently predictable in the initial training phase, while the other two outcomes were less predictable because each of them is preceded by cue C half of the time and cue D on the other half (Design 1). In the second phase, each outcome was fully predictable by a novel cue. We firstly observed the outcome predictability effect in the first experiment. The previously predictable outcome was more readily associated with a novel cue than the previously less predictable outcomes. However, this finding could not be reproduced in the following three experiments. All these results indicate that the finding from Experiment 1 is not replicable and this manipulation of outcome predictability in the present paradigm cannot reliably exert an effect on novel learning about this outcome. Based on the results of the first four experiments, we approached two other manipulations to differentiate the outcomes’ predictability (Design 2 and 3) in Phase 1 to detect if the different manipulations can influence the demonstration of the outcome predictability effect. Moreover, these two manipulations differ in the formation of contextual associations with outcomes. A stronger contextual association should be formed to the less predictable outcome with Design 3, but not with Design 2. In Phase 2, each outcome is, again, fully predictable by novel cues. Thus, through the comparison of data between two designs, we are able to determine whether context can mediate the demonstration of the outcome predictability effect. Experiment 5 to 7 examined these two designs. We observed once that the conduction of Design 2 demonstrated more rapid learning about the prior predictable than the prior less predictable outcome in Phase 2. However, through additional analyses we found that such a difference is based on a stronger association between context and the prior less predictable outcome showed in Phase 2, which is inconsistent with our expectation. Moreover, Experiment 7 with the additional instructional manipulations did not replicate this finding. Participants did not show a preference for the prior predictable outcome in Phase 2 learning, even though we explicitly informed them that the previous predictable outcome would also be predictable in Phase 2. Furthermore, the execution of Design 3 did not establish an effect of outcome predictability on subsequent learning, indicating that context cannot mediate the demonstration of the outcome predictability effect. Overall, our data suggest that the manipulation of the outcome’s predictability in the initial training phase cannot affect subsequent learning about the outcome in the present paradigm and processing of outcomes may differ from cues on the basis of predictability/predictiveness. We discussed that the different features between cues and outcomes might be responsible for the difference in cues and outcomes processing. Furthermore, we carefully reviewed the studies which have reported the outcome predictability effect and speculate that their findings might be paradigm-dependent