Abstract
Differential Evolution (DE) is a robust global optimization technique, yet its performance is highly dependent on the choice of mutation strategies and control parameters. This paper introduces ADE, a novel DE variant that incorporates a DE/lbest/1 mutation strategy and a hierarchical, two-level parameter adaptation mechanism. By utilizing an Indicator of Optimization State (IOS) to distinguish between exploration and exploitation phases, ADE dynamically adjusts parameters at both the population and individual levels. Evaluation on 33 benchmark functions and three real-world engineering problems shows that ADE provides superior robustness and scalability, particularly in complex, high-dimensional multimodal landscapes.
Introduction
Global optimization often involves navigating complex, multimodal landscapes where standard algorithms risk premature convergence. Differential Evolution (DE) is widely used due to its simplicity and effectiveness, but it remains sensitive to its primary control parameters: the scale factor ($F$) and crossover probability ($CR$). Traditional "greedy" strategies like DE/best/1 often lead to a loss of population diversity, causing the search to stall in local optima. While various adaptation schemes have been proposed, they typically focus on either the population as a whole or individuals in isolation. ADE addresses this by bridging the gap between global search dynamics and individual agent behavior.
Methodology
The ADE algorithm introduces two primary components to balance the exploration-exploitation trade-off.
1. DE/lbest/1 Mutation Strategy
To prevent the population from prematurely converging on a single global best, ADE employs a "local best" strategy. The population is partitioned into several non-overlapping groups. During mutation, each individual is guided by the best-performing member within its specific group rather than the global leader. This approach maintains multiple search fronts, preserving diversity throughout the optimization process.
2. Two-Level Parameter Adaptation
The core of ADE is its hierarchical adaptation scheme, which operates at two distinct levels:
- Population Level: The algorithm identifies the current "Optimization State" using the Indicator of Optimization State (IOS). This metric correlates an individual's fitness rank with its spatial distance rank from the global best. A high correlation suggests the population is clustered (exploitation), while a low correlation indicates a dispersed search (exploration).
- Individual Level: Based on the population-level state, parameters ($F_i, CR_i$) are fine-tuned for each individual. High-performing individuals near the best are tuned for local refinement, while underperforming individuals further away are encouraged to explore.
The detailed logic for the adaptation process, including the sorting and state-based adjustment of $F_p$ and $CR_p$, is shown below:
Experimental Evaluation
ADE was rigorously tested against four state-of-the-art DE variants (CoDE, JADE, jDE, and SaDE) using a comprehensive suite of 33 benchmark functions, including unimodal, multimodal, and hybrid composition functions from the CEC 2005 competition. Additionally, the algorithm was applied to three real-world problems from the CEC 2011 competition, involving sound wave parameter estimation and optimal control.
Results and Discussion
Experimental results highlight several key strengths of the ADE framework:
- Robustness: ADE achieved a 100% success rate on all unimodal and most multimodal benchmarks, outperforming competitors that occasionally failed on complex landscapes.
- Scalability: The performance gap between ADE and other variants widened as problem dimensionality increased to $D=100$, particularly on difficult multimodal functions.
- Real-World Utility: ADE demonstrated superior performance on the frequency-modulated sound wave problem (P1) and remained highly competitive on other engineering tasks.
- Efficiency: Despite the added complexity of sorting and distance calculations, ADE maintains the same asymptotic complexity as standard DE, $O(NP \cdot D \cdot G_{max})$.
Conclusion
ADE effectively manages the exploration-exploitation trade-off through its hierarchical parameter adaptation and local-best mutation strategy. By making the algorithm reactive to the real-time state of the search, ADE provides a more principled approach to parameter control than traditional ad-hoc methods. Future research will explore extending this two-level framework to other evolutionary algorithms and multi-objective optimization contexts.
Summary updated