Evaluating the Safety of Deep Reinforcement Learning Models using Semi-Formal Verification