Empirically validating a computational model of automatic behavior shaping

Abstract

Background Mobile sensing technology allows automated behavior shaping routines to be incorporated into health behavior interventions and other settings. In previous work, a computational model was built to investigate how to best arrange automatic behavior shaping procedures, but the degree to which this model reflects actual human behavior is not known. Purpose To translate a previously developed computational model of automatic behavior shaping into an experimental setting. Methods Participants (n = 54) operated a computer mouse and attempted to locate a hidden, randomly-placed target circle on a blank computer screen and clicks within some threshold distance of the target circle were reinforced by a pleasant auditory tone. As the trial progressed, the threshold distance narrowed according to a shaping function until eventually only clicks within the target circle were reinforced. Accumulated Area Under Trajectory Curves and Time Until 10 Consecutive Target Clicks were used to quantify the probability of the target behavior. Linear mixed effects models were used to assess differential outcomes for concave up, concave down, and linear shaping functions. Results In congruence with the computational model, concave-up functions most effectively shaped participants’ behavior, with linear and then concave-down shaping functions producing the next best outcomes. Conclusion Concave-up shaping routines most effectively generated target behavior, which should be confirmed in health behavior trials. The automatic shaping routines that this study helps develop can be applied in a number of domains, including exercise intensity and duration, tobacco/cannabis smoking, caloric intake, and screen time.