Action selection requires predicting future uncertainty

Abstract

Action selection based on predictive models involves look-ahead to estimate the likely outcome of an action sequence. However, prediction uncertainty grows with both the complexity of the model generating estimates and the length of look ahead. Thus, if humans utilize knowledge of the uncertainty associated with their estimates, changes in their response strategies should be observed as a function of the model complexity and predictive horizon imposed by different tasks. To test this hypothesis, we examined human performance in a predictive decision making task. Observers launched “arrows” toward targets in a computer display. They could adjust the arrow's position during its trajectory, and when they were certain the arrow was on target, they pressed the spacebar to relinquish control. Points were awarded as a function of the arrow's distance from the target when control was relinquished (missing the target resulted in zero points). Observers launched three arrow types in random blocks (identified by color) with dynamics models of increasing complexity (constant velocity, constant acceleration, constant jerk). The proper solution to this problem is to compute an expected value at each point in time and relinquish control when the expected value is at a maximum, which necessarily takes into account uncertainty. Observers' actual uncertainty was measured in interleaved blocks of trials in which the arrow's trajectory endpoint had to be extrapolated from a set trajectory. Behavior was consistent with the predictions of a model-based observer. Predictive uncertainty increased as a function of distance from the target and model-complexity as reflected in the timing of observers' relinquish control decisions, which were farther along the trajectory for the constant jerk model than constant acceleration than constant velocity. These results provide direct evidence that human observers do incorporate knowledge of uncertainty of look ahead when making decisions.

This research was supported by the Office of Naval Research grant N00014-07-1-0937.