Detection of change in nonstationary binary sequences

The present study was designed to investigate the ability of humans to detect changes in the underlying probability structure of binary sequences of events. Four Ss were presented with sequences of 30 events (zeros and ones) which had been generated by two underlying probability values. P₁ was the probability of a one in the first i events, P₂ was the probability of a one in the last 30-i events, and the distribution of i was uniform from 1 to 29. The Ss indicated, without feedback, where the transition point from P₁ to P₂ took place for four sets of problems in which P₁ and P₂ took on the following pairs of values: (.40, .60), (.20, .80), (.05, .60), and (.50, 1.00). Estimates by Ss were compared with true transition points, i, and with maximum likelihood estimates of i. The difference between P₁ and P₂ and the closeness of at least one P value to the boundaries, either zero or unity, interacted in determining the difficulty of the problem for Ss.