Using generative adversarial networks to evaluate robustness of reinforcement learning agents against uncertainties