Particle Swarm Optimization (PSO) has some shortcomings in solving complex multi-dimensional and multi-peak problems, such as low local search accuracy and easy to fall into local optimum. Therefore, this paper presents an improved particle swarm optimization algorithm based on average position learning. In this algorithm, neighboring particles with better adaptive values than the particles themselves are adopted as learning objects in the learning strategy,and the algorithm is divided into two stages with different updating speed formulas. In stage one, the average position of all particle positions in the whole population is introduced into the updating velocity formula. In the second stage, a new average position is introduced into the velocity updating formula, and the greedy strategy is adopted to select the individuals selected after each updating of the particles to be better than the historical optimal adaptive value of the population. In addition, the historical optimal positions of the corresponding individuals are stored, and their average positions are calculated after the end of the first stage. Taking the average position as the learning object can enhance the information exchange among particles, and balance the local development performance and global search ability of the algorithm. In the CEC2017 test function experiment, the experimental results show that the proposed algorithm has certain advantages compared with the other four algorithms.