Advancing Metaheuristic Optimization Through Bias Correction
The Sine Cosine Algorithm, commonly known as SCA, has served as a foundational population-based metaheuristic since its introduction in 2016. Researchers Hsing-Chih Tsai of National University of Kaohsiung and Jun-Yang Shi have now published a detailed examination titled True Sine Cosine Algorithm for Continuous Optimization Problems: Bias Analysis, Potential Corrections, and Fundamental Improvements. Their work appears in Applied Soft Computing, Volume 202, Part A, October 2026, article 115828. The full abstract and details are available at the original publication link: https://www.sciencedirect.com/science/article/abs/pii/S1568494626012767.
This study systematically identifies structural biases in the original SCA formulation and introduces corrected variants, most notably TSCA3. These changes address tendencies that favor solutions near the coordinate origin, improving reliability across shifted and high-dimensional problems. The research provides both theoretical analysis and extensive benchmarking on CEC2005 and CEC2019 suites, demonstrating clear performance gains.
Understanding the Sine Cosine Algorithm and Its Role in Optimization
Metaheuristic algorithms offer practical solutions for continuous optimization problems where traditional gradient-based methods struggle with non-convex, high-dimensional, or multimodal landscapes. SCA, proposed by Seyedali Mirjalili, updates candidate solutions using sine and cosine functions to oscillate around the current best position. This mechanism balances exploration in early iterations with exploitation later, making it attractive for engineering design, parameter tuning, and machine learning tasks.
The position update equation in the original SCA incorporates random parameters r1, r2, and r3 along with an absolute value operator. While simple to implement, this structure creates unintended search preferences. Tsai and Shi trace these preferences directly to specific components of the update rule, showing how they systematically direct agents toward the origin of the coordinate system.
Identifying Search Biases in the Original Formulation
Through careful mathematical analysis and targeted numerical tests, the authors demonstrate that SCA exhibits a pronounced bias toward problems whose global optima lie at or near the coordinate origin. When optima are shifted away from this point, performance degrades noticeably. The absolute value operator and the r3 parameter contribute to this central tendency by altering the effective search direction and step size in ways that favor the origin.
This bias mirrors issues previously identified in other algorithms such as the Gray Wolf Optimizer. It raises important questions about benchmark design, as many standard test functions place optima at the origin. The CEC2019 suite, for example, may inadvertently reward algorithms with this central search tendency, potentially overstating their general applicability.
Proposed Corrections: Removing Key Structural Elements
The researchers recommend two straightforward yet impactful corrections to the position update equation. First, removal of the absolute value notation eliminates an artificial constraint on movement direction. Second, elimination of the r3 parameter simplifies the search factor while preserving the core oscillatory behavior driven by r1 and r2.
These minimal modifications form the basis for a family of true SCA variants, collectively termed TSCA. The variants maintain the intuitive trigonometric foundation of the original algorithm while removing the sources of origin bias. Implementation remains straightforward, requiring only small changes to existing codebases.
Photo by Google DeepMind on Unsplash
Introducing TSCA3: Individual-Wise Updating and Evolutionary Processes
Building on the corrected foundation, TSCA3 incorporates individual-wise position updating and a lightweight evolutionary process. Rather than relying solely on the global best, each agent maintains and updates its own trajectory with controlled diversity. This addition helps regulate the exploration-exploitation balance across the population.
The evolutionary component introduces simple selection and replacement mechanisms that retain promising individuals without excessive computational overhead. Analyses of population diversity confirm that TSCA3 begins with broad exploration and transitions smoothly to focused exploitation in later stages, leading to faster and more reliable convergence on continuous optimization tasks.
Benchmarking Results on CEC2005 and CEC2019 Suites
Extensive experiments compare TSCA3 against the original SCA and several classical metaheuristics. On both low- and high-dimensional instances, TSCA3 achieves lower mean function errors and attains near-complete dominance in pairwise statistical comparisons. Performance advantages become especially pronounced on shifted problems and high-dimensional functions where origin bias previously hindered results.
Population diversity metrics and exploration-exploitation balance plots further validate the improvements. TSCA3 consistently reaches competitive or superior outcomes relative to established methods while retaining the simplicity that made SCA popular.
Implications for Benchmark Design and Algorithm Evaluation
The study underscores the need for greater diversity in optimal point locations within standard benchmark suites. Over-reliance on origin-centered functions can mask algorithmic weaknesses and inflate perceived performance. Researchers and practitioners are encouraged to incorporate shifted and rotated test cases when evaluating new or existing metaheuristics.
These findings also highlight opportunities for hybrid approaches that combine corrected SCA variants with other established techniques. The fundamental improvements in TSCA3 provide a cleaner starting point for such integrations across engineering, scientific computing, and data-driven applications.
Broader Impact on Continuous Optimization Research
Optimization algorithms underpin progress in fields ranging from structural engineering and controller design to feature selection in machine learning and resource allocation in large-scale systems. By correcting identifiable biases, TSCA3 and related variants offer more robust tools for these domains. The work encourages similar critical examinations of other popular metaheuristics to uncover hidden structural preferences.
Academics and research teams can build directly on the open suggestions for further algorithmic refinements. The emphasis on preserving core frameworks while making targeted corrections provides a model for incremental yet meaningful advances in the field.
Future Directions and Integration Potential
The authors note that TSCA3 serves as a robust alternative ready for practical deployment and further hybridization. Potential extensions include multi-objective formulations, constrained problem handling, and integration with machine learning pipelines. Continued development of diverse benchmark functions will support more reliable comparative studies.
Researchers interested in contributing to this area may explore applications in emerging domains such as neural architecture search or sustainable energy system optimization. The corrected trigonometric update mechanism offers a solid foundation for these explorations.
Practical Considerations for Researchers and Practitioners
Adopting TSCA3 requires only modest code adjustments from existing SCA implementations. The performance gains on challenging benchmarks suggest immediate benefits for many real-world continuous optimization tasks. Teams working with high-dimensional or shifted search spaces stand to gain the most from the bias corrections.
Documentation and reproducibility remain priorities; the authors provide clear descriptions of the algorithmic changes and experimental protocols. This transparency supports independent verification and extension by the broader optimization community.