Communication Load Balancing via Efficient Inverse Reinforcement Learning