Abstract: The particle swarm optimization(PSO) algorithm is a widely applied optimization algorithm in fields such as robot motion planning and signal processing. However, this algorithm is prone to getting stuck in local optima, resulting in premature convergence problem. One reason for this premature convergence problem is that the particle swarm relies solely on fitness values to select learning examples. To overcome this problem, a particle swarm optimization algorithm called FINPSO(particle swarm optimization algorithm based on a hybrid approach driven by fitness values, improvement rate, and novelty) is proposed. The algorithm introduces new metrics and utilizes a genetic algorithm to balance the exploration and exploitation of the population, reducing the likelihood of premature convergence in the particle swarm. Firstly, fitness values, improvement rate, and novelty are used as evaluation metrics for the particles. Each metric is independently employed to select learning examples, which are then stored in separate archives. During each velocity update, particles need to determine the weights of each metric and learn by selecting an example from each archive. Secondly, the algorithm incorporates a genetic algorithm for information exchange among particles. The genetic algorithm introduces cross-swapping and mutation, bringing more randomness to the population and enhancing its global search capability. Finally, eight variants of the PSO algorithm are used as comparative algorithms, and two CEC test suites are employed as benchmark functions in the experiments. The experimental results demonstrate that the FINPSO algorithm outperforms the existing PSO algorithm variants, reaching a state-of-the-art level.

Key words: particle swarm optimization, genetic algorithm, hybrid drive, global optimization algorithm, evolutionary algorithm

摘要： 粒子群优化（particle swarm optimization，PSO）算法是一种在机器人运动规划、信号处理等领域有广泛应用的优化算法。然而该算法易陷入局部最优解，从而导致早熟问题。出现早熟问题的原因之一是粒子群仅依靠适应度值选择学习范例。为了克服上述问题，提出了一种基于适应度值、改进率和新颖性混合驱动的PSO算法（particle swarm optimization algorithm based on hybrid driven by fitness values，improvement rate，and novelty，FINPSO）。在该算法中，引入的新指标和遗传算法会平衡种群的探索与开发，降低粒子群早熟的可能性。适应度值、改进率和新颖性会作为粒子的评价指标。各指标独立地选择学习范例并保存到不同的档案中。粒子每一次速度更新都要确定各个指标的权重，并从每个档案中选择一个范例学习。该算法采用了遗传算法进行粒子间的信息交流。遗传算法中的交叉互换和突变会给种群带来更多的随机性，提升种群的全局搜索能力。以八个PSO算法变体作为对比算法，两个CEC测试套件作为基准函数进行实验。实验结果表明，FINPSO算法优于已有的PSO算法变体达到最先进水平。

关键词: 粒子群优化, 遗传算法, 混合驱动, 全局优化算法, 进化算法