Q-learning based trajectory design for UAV communications network with fairly non-orthogonal multiple access