关键词: 海洋捕食者算法, Lloyd’s算法, Q-learning机制, 反向学习

Abstract: Large-scale optimization problems have been around across various domains in real life. However, these problems exhibit high-dimensional variables and complex interdependencies among variables, rendering traditional optimization algorithms often ineffective and inefficient. Given that swarm intelligence optimization algorithms possess strong global search capabilities, inherent potential for parallelism, and distributive characteristics, they are more suitable for addressing large-scale optimization problems. Nevertheless, such algorithms typically suffer from difficulties with balancing exploration and exploitation stages and are prone to local optima. Hence, it is imperative to investigate improvement strategies for swarm intelligence optimization algorithms. The Marine Predators Algorithm (MPA) is a novel swarm intelligence optimization search algorithm inspired by the foraging behavior of marine predators in nature. It features simplicity in principle, ease of implementation, and minimal parameter settings. However, similar to other swarm intelligence algorithms, the MPA also has its drawbacks, necessitating corresponding improvements to enhance its effectiveness.
To address the shortcomings of the MPA in achieving low solution accuracy and susceptibility to local optima when solving large-scale optimization problems, this study first utilizes the Lloyd’s algorithm to initialize the prey population. This ensures an even distribution of individuals throughout the solution space, thereby enhancing the population’s global search capability. Subsequently, the three position update strategies of the MPA algorithm are employed as actions, with the number of offspring individuals surpassing their parents as the state, and the reduction in the optimization objective value as the reward. By employing the Q-learning algorithm, the optimal position update strategy is determined for each iteration, thereby balancing the exploration and exploitation processes of the algorithm and preventing it from being trapped in local optima. Additionally, a reverse operation is applied to each optimized individual after each iteration, obtaining the reverse solution of the current population. This expands the search space of the IMPA algorithm, enhances the diversity of the population, and effectively prevents the algorithm from falling into local optima.
Finally, this paper conducts a comparative analysis of the IMPA with several existing improved versions of the MPA across 13 high-dimensional test functions. The analysis includes algorithm complexity analysis, convergence speed, and optimization performance analysis, as well as a statistical characterization based on Wilcoxon tests. The results indicate that the IMPA algorithm outperforms other comparative algorithms in terms of solution accuracy and convergence speed on high-dimensional problems. It demonstrates superior convergence capability and solution accuracy in large-scale optimization problems.
Currently, the discussion of the performance of the IMPA algorithm is based solely on experiments conducted on high-dimensional test functions without testing the algorithm’s performance on actual large-scale optimization problems. The next step in research primarily involves applying the IMPA algorithm to specific large-scale constrained optimization problems and practical engineering issues.

Key words: Marine Predators Algorithm, Lloyd’s algorithm, Q-learning mechanism, opposition-based learning