Abstract
The Lévy flight Shuffle Frog Leaping Algorithm (LSFLA) is a variant of the Shuffled Frog Leaping Algorithm (SFLA) that significantly enhances performance. However, it still suffers from defects such as poor convergence and low efficiency. This paper proposes an improved version, the LSFLA based on Differential perturbation and Quasi-Newton search (DQLSFLA). DQLSFLA introduces an all-solution updating method to replace the traditional worst-solution update, reducing computational complexity and omitting redundant sorting steps. Furthermore, it employs a hybrid subgroup strategy combining random differential perturbation and an improved Lévy flight method to balance global exploration and local exploitation. Finally, a quasi-Newton local search is integrated to accelerate convergence in the final stages. Experimental results on the CEC-2014 test suite and real-world applications, including medical image enhancement and the Quadratic Assignment Problem (QAP), demonstrate that DQLSFLA outperforms LSFLA and other state-of-the-art algorithms in both optimization quality and efficiency.
Introduction
Intelligent Optimization Algorithms (IOAs) are essential tools in artificial intelligence, designed to solve complex, non-linear problems by mimicking natural processes. The Shuffled Frog Leaping Algorithm (SFLA) is a prominent example, simulating a group of frogs searching for the best food source. While effective, SFLA and its variants like LSFLA often struggle with low efficiency due to frequent sorting, premature convergence into local optima, and high sensitivity to parameter settings.
The DQLSFLA was developed to address these bottlenecks, creating a more robust optimizer capable of handling high-dimensional challenges in fields ranging from logistics to medical imaging.
Methodology: The DQLSFLA Framework
The DQLSFLA introduces three primary innovations to the standard frog-leaping paradigm to enhance its search capabilities and speed.
1. All-Solution Updating
Traditional SFLA variants typically update only the "worst" frog in a subgroup during each iteration. This requires the algorithm to re-sort the population constantly to identify the worst performer, a computationally expensive process. DQLSFLA replaces this with an all-solution updating mechanism. By updating every frog in a subgroup simultaneously, the algorithm eliminates the sorting step at every sub-iteration, reduces overhead, and facilitates parallel computing architectures.
2. Hybrid Subgroup Strategy
To maintain a balance between "exploration" (searching new areas) and "exploitation" (refining known solutions), DQLSFLA splits each subgroup into two halves:
- First Half: Utilizes a Random Differential Perturbation approach. This maintains population diversity and prevents the algorithm from getting trapped in local optima.
- Second Half: Employs an Improved Lévy Flight method. This uses a combination of small steps and occasional long jumps, allowing the frogs to navigate the search space more effectively than standard random walks.
3. Quasi-Newton Local Search
Bio-inspired algorithms are excellent at finding the general vicinity of a global optimum but often lack the precision to find the exact peak. DQLSFLA addresses this by triggering a Quasi-Newton local search during the final 3% of the evaluation budget. This mathematical "finishing move" provides the high-precision refinement necessary for complex engineering problems.
Experimental Evaluation and Results
The algorithm was rigorously tested against three distinct benchmarks to verify its versatility and power.
CEC-2014 Benchmark Performance
On the CEC-2014 test suite, which includes 30 complex mathematical functions, DQLSFLA achieved the best average rank (2.43) among 11 state-of-the-art algorithms. Notably, in 50-dimensional tests, DQLSFLA completed its optimization in 22.1 seconds, a 31% reduction in runtime compared to the 32.2 seconds required by the baseline LSFLA.
Real-World Applications
- Medical Image Enhancement: When applied to chest bone fluoroscopic and blood vessel images, DQLSFLA produced higher objective evaluation scores ($F(Z)$) than LSFLA, resulting in significantly clearer textures and better contrast.
- Quadratic Assignment Problem (QAP): In logistics optimization (assigning 20 facilities to 40 locations), DQLSFLA found lower-cost solutions with higher stability and lower standard deviation than competing methods.
Conclusion
DQLSFLA represents a significant step forward in bio-inspired optimization. By streamlining the update process and combining stochastic search with rigorous local refinement, it achieves a superior balance of speed and accuracy. Future research will focus on making the transition to the Quasi-Newton phase more adaptive and applying the algorithm to broader engineering design and resource scheduling challenges.
Summary updated