the value of generative adversarial training for model-based reinforcement learning (RL) with offline data, especially