Abstract
Evolutionary algorithms are powerful tools for global optimization but often suffer from slow convergence in the final stages of search. This paper introduces HyGO (Hybrid Genetic Optimisation), a framework that combines the global search capabilities of Genetic Algorithms (GA) and Genetic Programming (GP) with the local refinement precision of the Downhill Simplex Method (DSM). A key innovation is a "degeneracy-proof" DSM that prevents the local search from stalling in high-dimensional spaces. HyGO demonstrates superior performance across analytical benchmarks and complex engineering tasks, notably achieving a 20.56% drag reduction in automotive aerodynamic simulations.
Introduction
In engineering design, optimization tasks generally fall into two categories: parametric learning (optimizing numerical values) and functional learning (discovering mathematical rules or control laws). While evolutionary algorithms like GAs are robust against local optima, they are often inefficient at "polishing" a solution once the global neighborhood is found. Conversely, local search methods like the Downhill Simplex Method (DSM) are fast but prone to getting stuck or failing due to simplex "degeneracy"—where the search volume collapses into a lower-dimensional subspace. HyGO is designed to bridge this gap, providing a robust, two-stage hybrid approach.
The HyGO Framework
HyGO employs a two-stage, alternating strategy to balance global exploration and local exploitation.
- Exploration Phase: This phase uses either a Genetic Algorithm (for parameters) or Linear Genetic Programming (for functions) to sample the search space broadly. It identifies promising regions and maintains population diversity.
- Exploitation Phase: The best individuals from the exploration phase are used to initialize a local search using the Downhill Simplex Method (or its variant, Subplex).
To address the fragility of local search, HyGO implements a degeneracy-proof DSM. By monitoring the "flatness" of the simplex using a coefficient of determination ($R^2$), the algorithm can detect when the simplex is collapsing. If degeneracy is detected, HyGO mathematically injects a new point orthogonal to the collapsed plane, "re-inflating" the search volume and allowing the optimizer to continue its descent.
Genetic Mechanisms for Exploration
The exploration phase is tailored to the type of problem being solved. For parametric optimization, a standard Genetic Algorithm is used. For functional learning, Linear Genetic Programming (LGP) is employed, which evolves sequences of instructions that manipulate registers to form mathematical expressions.
As shown in the figure above, both methods utilize crossover and mutation operators to evolve the population. LGP is particularly valuable for engineering because it produces interpretable symbolic expressions rather than "black box" models.
Experimental Validation and Results
The framework was validated across three tiers of difficulty:
- Analytical Benchmarks: 15 classic functions (e.g., Rosenbrock, Ackley) in both 2D and 25D. HyGO reached optima in significantly fewer evaluations than standard GAs (e.g., ~265 vs ~2,542 for the Booth function).
- Functional Learning (Landau Oscillator): HyGO successfully evolved a simple, elegant control law—$7.262 \cdot (\sin(a_1) + a_1 - a_2)$—to stabilize nonlinear fluid dynamics.
- Aerodynamic Optimization (Ahmed Body): In a high-stakes automotive benchmark, HyGO achieved a 20.56% reduction in drag, outperforming standard GAs (10.88%) and state-of-the-art gradient-based methods (17.49%).
Conclusion
HyGO effectively addresses the "last mile" problem of evolutionary optimization by hybridizing global search with a robust local refinement stage. Its ability to handle both parametric and functional learning makes it a versatile tool for complex engineering challenges. Future developments will focus on automated hyperparameter tuning and adaptive switching between phases to further enhance its "plug-and-play" capabilities.
Code Availability: The HyGO framework is available as an open-source Python package on PyPI and GitHub.
Summary updated