TL;DR
This paper introduces an improved version of the Educational Competition Optimizer (ECO), called IECO-MCO, by incorporating three novel "multi-covariance learning operators." These operators aim to better balance exploration and exploitation, leading to faster convergence, improved stability, and a reduced tendency to get stuck in local optima, outperforming the original ECO and other state-of-the-art algorithms on benchmark and engineering design problems.
What Problem Does This Paper Address?
Many real-world problems, especially in engineering and scientific research, are complex optimization problems characterized by non-linear dynamics, non-convex structures, and high-dimensional spaces. While metaheuristic algorithms are powerful tools for tackling these challenges, existing ones often struggle with a critical trade-off: balancing exploration (searching broadly across the solution space) and exploitation (intensively searching promising regions). This imbalance can lead to premature convergence to suboptimal solutions (local optima) and slow convergence speeds. The original Educational Competition Optimizer (ECO), despite its promise, suffers from these limitations, particularly when dealing with highly complex optimization tasks. This paper aims to enhance ECO's ability to navigate these complex landscapes more effectively.
Methods
The core innovation of this paper is the introduction of three novel covariance learning operators integrated into the ECO framework, creating the IECO-MCO algorithm. These operators replace the original position update mechanisms for "school agents" in ECO across its three stages (primary, middle, and high school):
- Gaussian Covariance Operator: Used in the primary school stage, this operator leverages information from high-performing agents to guide the population towards advantageous directions, promoting broader exploration and improving overall population quality.
- Shift Covariance Operator: Applied in the middle school stage, this operator balances exploration and exploitation by adjusting agent movement directions based on multiple reference points, ensuring a more nuanced search.
- Differential Covariance Operator: Employed in the high school stage, this operator enhances population diversity and helps escape local optima by utilizing the differences between random solutions and the best/worst individuals.
These operators are designed to work synergistically, improving the algorithm's ability to explore the search space, exploit promising areas, and maintain population diversity.
Datasets
The paper evaluates IECO-MCO using two standard benchmark test suites:
- CEC 2017: Consists of 29 functions covering unimodal, multimodal, hybrid, and composition types.
- CEC 2022: Consists of 12 functions.
Additionally, the algorithm's practical applicability is tested on ten real-world constrained engineering design optimization problems, including spring design, pressure vessel design, and truss design problems.
Key Insights and Ideas
- The power of covariance: The paper demonstrates that by intelligently modeling and utilizing covariance information (how variables change together), the search process can be significantly guided towards better solutions.
- Stage-specific optimization: Applying different covariance operators at different "educational stages" (primary, middle, high school) allows for a tailored approach to exploration and exploitation, mimicking a learning progression.
- Synergy over isolation: While individual covariance operators show improvement over the basic ECO, their combined use in IECO-MCO yields the best results, highlighting their complementary nature.
- Robustness across dimensions: IECO-MCO maintains stable performance across varying problem dimensions, suggesting good scalability.
Key Results
- Superior Performance: IECO-MCO consistently outperformed the basic ECO, its own variants (GECO, SECO, DECO), five basic algorithms (SAO, CFOA, DBO, QIO, AE), and five improved algorithms (RDGMVO, ISGTOA, AFDBARO, MTVSCA, ALSHADE) across the CEC 2017 and CEC 2022 benchmark suites.
- Ranking: IECO-MCO achieved an average Friedman rank of 1.93 on CEC 2017 and 1.79 on CEC 2022, significantly better than the basic ECO's ranks of 4.26 and 4.63, respectively.
- Convergence and Stability: Statistical tests (Friedman, Kruskal-Wallis, Wilcoxon rank-sum) confirmed IECO-MCO's superior convergence speed, stability, and ability to avoid local optima.
- Engineering Applications: The algorithm successfully solved ten real-world constrained engineering design problems, demonstrating its practical effectiveness.
Limitations and Future Work
- Suboptimal performance on specific functions: While IECO-MCO excels overall, the authors note that its suboptimal performance on certain functions motivates further improvements.
- Future Directions:
- Application Expansion: Applying IECO-MCO to more complex domains like mission planning, medical image segmentation, and automated artistic design.
- Multi-objective and Binary Versions: Developing multi-objective versions to handle conflicting goals and binary versions for discrete optimization problems.
- Integration with AI: Exploring the integration of IECO-MCO with reinforcement learning and deep learning techniques for adaptive strategy selection.
- Generalization of Operators: Investigating the applicability of the proposed covariance operators to other metaheuristic algorithms.
Summary updated