In order to address the problems that the original teaching-learning-based optimization (TLBO) cannot balance the global exploration and the local exploitation well, and the current opposition-based learning (OBL) strategies are too simple, a multi-opposition teaching-learning-based optimization (MOTLBO) was proposed by combining various opposition-based strategies with TLBO. Firstly, learning from the idea of OBL, a nonlinear mixed opposition-based learning model with a Sigmoid function and gradual evolution change was designed, in which the boundary search information and the population historical search information systematically were taken into account. Secondly, a self-learning stage guided by search boundary was added based on the existing teaching and learning stages, which can enhance the population diversity greatly. Finally, combining the mixed model with each stage of the algorithm, the calculation methods of opposition-based solutions according to mean individual, randomly selected individual, and optimal individual, respectively, were presented, which fully utilized the historical search experience of the population and hence improved the convergence precision and speed. The Benchmark test functions with different characteristics were used to test the nonlinear mixed opposition-based learning model and convergence performance of the proposed algorithm. The experimental results showed that the nonlinear mixed opposition-based learning model has stronger capability of global exploration and local exploitation than the one that only uses the boundary search information or population historical search information. By compared with the TLBO and some of its current outstanding variants, the proposed algorithm obtained high optimization accuracy and stability while guaranteeing the convergence speed, achieving great improvements on comprehensive performances. Moreover, the experimental results of the spread spectrum radar Polly phase code design demonstrated that the proposed algorithm can escape from local optimum effectively. Hence, the proposed algorithm can also be used to handle the real-word engineering optimization problems.