The Study About The Process Of Probability Learning

Probability matching is a classical behavioral anomaly observed in lab and field studies.When asked to make predictions of two mutually exclusive random events from Bernoulli process repeatedly,people tend to match their predictive frequences to observed event frequences,which violate maximization assumption in rational decision theory.Previous studies suggest there are three independent psychological processes that lead to this behavior: associative learning,pattern searching,and cognitive heuristics.Associative learning refers to a learning process of extrapolating short-term frequency experience;pattern searching is a more sophisticated strategy which exploit sequence patterns that may exist in the random process;heuristic is a cognitive shortcut to diversify responses and balance proportion of each type of events in the predictive sequence.Previous researchers attempt to distinguish these three hypotheses by manipulating the cognitive load and cognitive ability,but due to confounding of latent variables,the conclusions are inconsistent.On the basis of the predecessors,our study tried to use a cognitive auxiliary procedure,a dynamic feedback window,to extend short-term memory capacity.By manipulating the content and length of this window,we designed three experiments to separate cognitive process of probability learning.We showed that the three hypothesis of probability matching will make independent prediction of how behaviors will be sensitive to these manipulations.Experiment 1 manipulate the content of the window in separate conditions: update recent events,update recent selections,no feedback.Compared with the control condition,update selections reduced behavioral sensitivity of response to recent win/loss experience;on the country,update events made to predictive responses more sensitive to sequential patterns.The former may reflect an exploration behavior independent of recent event,while the latter may suggest a change of tendency to searching patterns.Further analysis of the learning curve shows that for a part of the subjects,event-sensitive responses(i.e.win-stay-loss-switch)increases gradually with the repetition of trials.While the others showed an increasing tendency to adopt a single optimal response strategy.These results suggest that associative learning and pattern searching,the two main processes of probability matching behavior,may coexist in the repeated prediction task.Experiment 2 manipulated window length and controlled the content.Recent events are updated with two level of length,3 and 9.The results further support the conclusion of experiment 1.First,by checking how responses are sensitive to streaks/runs of events,we find that the recency curve showed a complex wavy form: as the length of the streak increases,the tendency to follow the last event(recency effect)increase when the streak is short(positive recency effect at 1-3 repeats),decreases at medial range of length(negative at 4-7),increase again when the length is long enough(e.g.positive again after 7).As subjects gained experiences,negative part of the curve is gradually replaced by the positive effect,suggesting that the process of pattern searching is overturned a process of associative learning process.Secondly,the pattern of recency curve also changed when window length is increased(from 3 to 9),: at the high-probability-event end of the curve,the shifting from decreasing to increasing occurred early in the condition of 3,but latter in the condition of 9;but at the low-probability-event end,the opposite is true.This result implies that the process of searching patterns may embedded in the strategies of encoding and inferencing of streaks of events,which are sensitive to task situations.Finally,there is a positive correlation between the tendency of pattern searching reported in the questionnaire and the probability matching behaviors showed in the repeated prediction task.These are further evidence support support for the hypothesis of pattern searching.In the 3rd experiment,the effect of feedback on learning was studied.A new form of feedback is adopted to update the recent selection,but distinguished win trials from the losses,in the dynamic window.Although the experiment did not find that this form of feedback could accelerate learning of the optimization strategy,it did a first attempt to utilizing dynamic feedback to facilitate learning,provided a new approach and refine of feedback design in learning experiment.In conclusion,by using a dynamic feedback design to manipulate short-term memory,our studies separated cognitive process during probability learning,provided new evidence to associative learning and pattern searching hypothesis of probability learning.These are both meaningful from theoretical and methodological views.