“Don''t think too much!” — Artificial somatic markers for action selection

摘要

Humans could not effectively perform everyday decisions if they could only count on purely rational evaluations of the available response options. According to the Somatic Marker hypothesis, before such evaluations take place, emotional signals drastically reduce the space of available options and provide essential biases for decision making. Within the context of action selection, we present a model of the somatic markers' mechanism, encompassing three related processes of an AI agent: (1) Creating markers from the agent's experiences; (2) Combining markers and producing generalizations; (3) Using existing markers to aid in action selection. Computer simulations in two different domains respectively show: (a) a resulting behavior that is similar to that obtained from an experiment with human subjects, and (b) an intuitive creation and effect of generalized markers.