Learning Models of Adversarial Agent Behavior under Partial Observability