Source
Howes, Duggan, Kalidindi, Tseng and Lewis (2016), LIT-1062 (Cognitive Science 40(5), 1192–1223), read in NOTE-882: both experiments, the choice/no-choice analysis and the strategy-change result of Experiment 2.
What was actually shown
The task. A mock email shows one appointment at a time (“09.00: Appointment with NAME”); after a trial’s list, the participant switches to a calendar and types the names in order. “Participants could not go back to correct errors”, and nothing lets them return to the email during recall. The only strategy dimension is list length, 3 to 9. A shorter list means more trials, each with its overhead, to reach a fixed total. Reward rate is points per trial duration, including the move to the next trial.
The design. A no-choice phase ran every participant through every length in random order; a choice phase then let them pick. Each person’s boundedly optimal length is the argmax of their own mean rate. “the cost of optimization is paid by the analyst, not by the participant”.
Experiment 1 (N = 40). The person’s optimum predicted 55% of choices, one fewer 8%, one more 17% (optimum beat each, p < .001). Its correlation with mean chosen length was r = .35. Error cost (each wrong item scores zero, or one error voids the trial) did not change chosen length (5.05 against 5.00).
Experiment 2 (N = 20). An exponential points schedule raised chosen length (6.99 against 5.07, d = 1.33). The optimum predicted 46% of choices, neighbours 12% and 14%; correlation r = .77. People also changed strategy outside the space: on 49% of exponential trials they typed first letters in every box before completing names, and chose longer lists when they did.
Departures. Across both experiments, 13 of 60 “persistently selected a list length, and therefore a strategy, that was not bounded optimal”, 10 of them by one item more. In Experiment 1, eight became less likely to choose their optimum as trials went on.
What could have come out otherwise. Choices could have matched each strategy’s probability of being best, or sat one item below the optimum, as an offloading preference would predict. Neither happened.
What this does not say
- That this is a choice between memory and the world. There is no look-up. Lieder and Griffiths (LIT-1071, NOTE-873) call it a memorize-or-look-up trade; the task gives no way to consult the source again.
- That a model predicts the choice. No capacity, look-up time or error parameter appears. The prediction is each person’s measured best, and the bounds enter only through it.
- That the optimum was computed from data independent of the choices. Which trials enter each person’s mean rate is not stated precisely, and the paper’s distributions pool both phases.
- That people price errors. Voiding the whole trial on one error changed nothing; what moved choice was the payoff for long lists.
- That choice is optimal in general. About a fifth chose otherwise, persistently, and the paper’s explanations for this are untested.
- Anything about precomputation, reuse, staleness or search. Items are used once, within seconds; nothing stored serves a later trial.
Connections
- QUESTION-095: the “defer and re-read” strategy with the re-read removed. Its one addition to the question’s terms is a price on a wrong stored item, which did not move behaviour. Practice made a chosen length better, the mirror of the erosion of an unused route in LIT-1059.
- THEORY-346 fits group means with a time-only model; this tests, per person, whether people find the best point of their own trade, and keeps the earlier conclusion that they neither minimize memory nor favour the world.