Title: Assessing the Roles of Central, Peripheral, and Temporal Vision in Human Decision-Making During Atari Gameplay

Abstract: This research investigates the extent to which various visual data streams influence human decision-making processes within dynamic visual settings. Leveraging Atari-HEAD, a comprehensive dataset featuring synchronized eye-tracking and Atari gameplay, we propose a controlled ablation framework designed to deconstruct and quantify the impact of peripheral visual input, explicit gaze data represented by gaze maps, and historical behavioral states. We evaluate action-prediction networks across six distinct configurations, each manipulating the availability of these specific information sources. Our findings, drawn from 20 different games, reveal that peripheral vision is the dominant factor; removing this input results in the most significant decline in prediction accuracy, with median drops ranging from 35.27% to 43.90%. In contrast, the exclusion of gaze information leads to smaller accuracy reductions of 2.11% to 2.76%, while the removal of past-state information causes a wider variance in performance loss, between 1.52% and 15.51%. The higher end of this latter range is likely driven by the elimination of peripheral information leakage. To supplement these aggregate accuracy metrics, we employed clustering techniques on states based on true-action probabilities generated by the various model setups. This approach delineates distinct behavioral patterns, such as focus-centric, periphery-centric, and context-dependent decision-making scenarios. Collectively, these outcomes indicate that human choices in Atari are heavily reliant on visual cues outside the immediate gaze focus, and the introduced framework offers a viable method for estimating the influence of such information sources through behavioral analysis.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC