Abstract: Evolutionary Algorithms (EAs) are highly effective for global optimization in complex design spaces but are often limited by the high computational cost of fitness evaluations. This study evaluates the integration of neural networks as surrogate models to accelerate the optimization process while avoiding "false convergence." We compare four individual-based evolution control strategies—Best Strategy, Pre-Selection, Clustering Technique, and a hybrid approach—across three benchmark functions: Sphere, Rosenbrock, and Ackley. Our findings suggest that the Pre-Selection strategy provides the most consistent performance gains, and that dynamically optimizing the neural network's architecture further enhances search efficiency.
Introduction
In fields such as aerospace and automotive engineering, calculating the "fitness" of a design—like the aerodynamic drag of a turbine blade—often requires computationally intensive simulations. While Evolutionary Algorithms (EAs) are excellent at navigating these complex spaces, their reliance on a high volume of evaluations makes them expensive to deploy.
To mitigate this, researchers use meta-models (surrogates), such as neural networks, to provide fast mathematical approximations of fitness. However, relying exclusively on a model can lead to "false convergence," where the EA optimizes for errors in the surrogate rather than the actual physics. The core challenge is Evolution Control: determining how to intelligently balance a small number of expensive, high-fidelity evaluations with many cheap, low-fidelity estimates.
Evolution Control Framework
The authors investigated four "individual-based" evolution control frameworks designed to manage surrogate usage:
- Best Strategy (BS): Offspring are first evaluated by the neural network; only the top-performing individuals are then re-evaluated using the real fitness function.
- Pre-Selection (PreS): A large offspring pool is generated, the neural network filters for the most promising candidates, and only these are evaluated with the real function.
- Clustering Technique (CT): Offspring are grouped into clusters based on features, and only the individual closest to each cluster center receives a real evaluation.
- Clustering Technique with Best Strategy (CTBS): A hybrid where the predicted "best" individual within each cluster is selected for real evaluation.
A key innovation in this study is Structure Optimization. Rather than using a static model, the authors employed a genetic algorithm to evolve the neural network's architecture (adding or removing neurons and connections) during the optimization process. This ensures the surrogate adapts its learning capacity as the search moves into different regions of the design space.
Experimental Results
The strategies were tested on three 10-dimensional benchmark functions: the unimodal Sphere, the narrow and curved Rosenbrock valley, and the multi-modal Ackley landscape.
Key Findings
- Stability of Pre-Selection: The Pre-Selection (PreS) strategy emerged as the most reliable. It consistently outperformed the standard Evolution Strategy (PlainES) across different runs, particularly on the complex Ackley landscape where it showed significantly lower variance than other methods.
- Adaptive Modeling: Static networks struggled with the Rosenbrock function's complexity. Allowing the neural network structure to evolve on the fly helped the system adapt to local landscape features, though this function remained the most challenging for all methods.
- Efficiency Gains: All model-assisted strategies significantly reduced the number of expensive evaluations required to reach high-quality solutions. On the Sphere function, the improvement over PlainES was substantial across the board.
Conclusion and Future Work
The study confirms that individual-based evolution control is a powerful tool for reducing the computational burden of EAs. The most successful strategies use the surrogate model to filter candidates but rely on high-fidelity evaluations to select the parents for the next generation, providing a safeguard against false convergence.
Future research will explore the use of ensembles of neural networks to improve uncertainty estimation. Additionally, making the evolution control parameters adaptive—allowing the algorithm to automatically shift between model-heavy and evaluation-heavy modes based on surrogate accuracy—remains a promising avenue for further optimization.
Summary updated