In recent years, individual learning path planning has become prevalent in online learning systems, while few studies have focused on teaching path planning for traditional classroom teaching. This paper proposes a target-oriented teaching path optimization scheme for cloud computing-assisted instructions, in which a sequence of learning contents is arranged to ensure the maximum benefit for a given group of students. First, to evaluate the teaching performance, we investigate various student models and define some teaching objectives, including the pass rate, the excellence rate, the average score, and related constraints. Second, a new Deep Reinforcement Learning (DRL)-based teaching path planning method is proposed to tackle the learning path by maximizing a multi-objective target while satisfying all teaching constraints. It adopts a Proximal Policy Optimization (PPO) framework to find a model-free solution for achieving fast convergence and better optimality. Finally, extensive simulations with a variety of commonly used teaching methods show that our scheme provides nice performance and versatility over commonly used student models.