Improving deep reinforcement learning by safety guarding model via hazardous experience planning