Predicting Surprise Judgments from Explanation Graphs


Surprise is a ubiquitous phenomenon that is implicated in many areas of cognition, from learning, to decision making, to creativity. For example, it has recently been proposed as a trigger for learning in robotic agent architectures. This paper describes a novel cognitive model of surprise based on the idea that surprise is fundamentally about explaining why the surprising event occurred; events that can be explained easily are less surprising than those that are more difficult to explain. Using explanations that people have produced, this surprise model builds a directed graph of explanations that link the setting and outcome of a given scenario, and uses this graph to predict surprise ratings. Simulations are reported which show that the model’s performance corresponds closely to the psychological evidence, as measured by people’s ratings of different surprising scenarios.