Neural structure mapping in human probabilistic reward learning