Simulating and classifying behavior in adversarial environments based on action-state traces: an application to money laundering