Stochastic operator models for multiple target search tasks

A novel approach to stochastic operator modeling for multiple target search tasks Is introduced. Unlike the large majority of experiments in literature the focus Is on modeling and prediction of the sequence of detected and identified targets, rather than the prediction of time consumption. We investigate the prediction accuracy of three stochastic operator models: (1) a "classic" Markov Chain of first order, (2) a factorial Hidden Markov Model, and (3) a Variable-Length Markov Chain, The parameters are estimated with the help of empirically acquired traces from search experiments with 37 experienced operators. Five search scenarios with a varying number of targets were studied. Moreover, the prediction accuracy of two deterministic, but semantic models are Investigated: (4) a nearest-neighbor estimator and (5) a nearest-neighbor estimator with perfect procedural memory. Both semantic models do not rely on the empirical data, but Integrate task-specific a-priori knowledge. The results show a significantly highest prediction accuracy of the nearest-neighbor estimator with perfect memory, while the memoryless estimator has the lowest accuracy. These findings are Independent of the number of targets and Indicate the remarkable ability of the operator to memorize already attended targets. The Variable-Length Markov Chain has the significantly highest prediction accuracy among the stochastic models. Finally, a regression formula of the prediction accuracy depending on the number of targets is derived.