A Transfer Approach Using Graph Neural Networks in Deep Reinforcement Learning