Marine Predators Algorithm (MPA) is a newly nature-inspired meta-heuristic algorithm, which is proposed based on the Lévy flight and Brownian motion of ocean predators. Since the MPA proposed, it has been successfully applied in many fields. However, it exists several shortcomings, such as falling into local optimum easily and precocious convergence. To balance the exploitation and exploration ability of MPA, a modified marine predators algorithm hybridized with teaching-learning mechanism is proposed in this paper, namely MTLMPA. Compared to MPA, the proposed MTLMPA has two highlights. Firstly, a kind of teaching mechanism is introduced in the first phase of MPA to improve the global searching ability. Secondly, a novel learning mechanism is introduced in the third phase of MPA to enhance the chance encounter rate between predator and prey and avoid premature convergence. MTLMPA is verified with 23 benchmark numerical testing functions and CEC-2017 functions. Experiment results reveal that the MTLMPA is competitive compared to the state-of-art heuristic optimization algorithms.