Proceedings of the Annual Meeting of the Cognitive Science Society

Executive functions support flexible, goal-directed behavior in dynamic environments. Existing models typically attribute the substantial variability in human performance on such tasks to abstract, resource-based accounts of attention and working memory, leaving the nature of the underlying resource underspecified. Here, we investigate executive-function variability using a modified, continuous version of the Pong game that imposes sustained demands on multi-object tracking, control, and action selection. To account for the broad distribution of human performance levels, we compared a performance-optimized deep reinforcement learning agent with a biologically plausible spiking neural network that incorporates explicit neural resource constraints. Whereas the unconstrained artificial agent achieved superhuman performance, the spiking model reproduced the full spectrum of human scores. Systematic variation in neural resources, perceptual noise, and processing delays produced performance regimes corresponding to low, median, and peak human behavior, including saturation at object counts consistent with known working-memory limits.