A model of symbolic processing in Raven's progressive matrices

Raven's progressive matrices probe the capacity to discover a common rule in a 3 by 3 matrix of stimuli by requiring the correct selection of a probe stimulus for one stimulus location intentionally left blank. This task is purely visual and excludes verbal cues. We propose a model of the performance of this task that classifies a variety of visual stimuli, either as rotated, scaled, or with elements added or subtracted or otherwise altered within each of the 3 by 3 matrix's fields. In our model, Q-learning uses a state space based on the transitions between fields of the 3 by 3 matrix of stimuli. Actions are modeled as alterations of visual stimuli between these matrix's fields. Using Q-learning our model successfully selects the correct probe stimulus from eight choices offered in the task.