This paper aims to focus on the state-of-the art optimization variants of particle swarm optimization algorithm.
A state evaluation method based on two factors is introduced in the MIM-PSO algorithm to monitor the evolutionary state of the population in real-time. Based on four evolutionary states (exploration, exploration, development, and convergence), four different learning strategies were adopted, namely: random particle optimal position learning strategy, center position adaptive learning strategy, multi elite dimension selection strategy, and distance based local position search strategy. In addition, a conditional restart strategy is adopted to help the population escape from local optima. And applied to the hyperparameter optimization of neural network models for performance testing.
The algorithm was tested on different dimensional test functions of CEC2017 and CEC2021, and tested on standard databases and engineering applications. The results indicate that MIM-PSO has superior optimization performance, which can balance learning strategies and evolutionary states, and more efficiently find the global optimal solution.
The MIM-PSO proposed in this study is practical and feasible in solving complex and high-dimensional problems.
