When designing a system to transmit satellite images, it is important to balance the need for a high compression rate with the ability to accurately reconstruct them. The discrete wavelet transform (DWT) can be used to compress images and can be fine-tuned for specific classes of images, but it requires substantial hardware resources to perform floating-point multiplications. Separate research efforts to evolve wavelets and hardware multipliers have been performed, but not together. This work explores the use of several evolutionary and coevolutionary algorithms to address this gap, focusing on different ways of combining wavelets and hardware. The first algorithm evolves them in two separate stages, a second evolves them in pairs, and the third algorithm cooperatively coevolves two populations of wavelets and hardware. The results indicate that problem decomposition is most beneficial when evolution allows individual components to be optimized separately. Additionally, our evolved wavelets show improved accuracy over existing wavelets at high compression ratios, and the evolved hardware uses four times fewer components as a human-designed approach. This work demonstrates strong potential for effective image compression and, more broadly, the use of coevolutionary algorithms to fine-tune circuit design alongside secondary systems defining the needs of those circuits.
Electroencephalography (EEG) data is highly sensitive, requiring strong privacy protection in brain–computer interfaces and neurological diagnosis. Homomorphic encryption enables secure EEG classification with deep neural networks, but standard activation functions (e.g., ReLU) are incompatible with encrypted arithmetic, causing severe computational overhead. Polynomial approximations mitigate this issue yet fail to capture EEG’s distinctive spectral patterns, temporal dynamics, and inter-subject variability. We propose a surrogate-assisted genetic algorithm (SA-GA) to optimize activation function approximations for encrypted deep learning models. SA-GA uses fixed-length chromosome encoding and a Gaussian Process surrogate to efficiently search the approximation space while avoiding costly encrypted evaluations. Applied to motor imagery classification (BCI Competition IV Dataset 2a), our method achieves 89.7
Single Image Super-Resolution reconstructs high-fidelity images from low-resolution inputs under tight computational budgets, essential in fields like medical diagnostics, remote sensing, surveillance, and aerospace imaging. To ease manual network design, Neural Architecture Search is a key AutoML tool, though its effectiveness hinges on often costly and imperfect performance evaluations. This paper compares two efficient predictors: (i) SynFlow, a zero-cost method, and (ii) an XGBoost regressor, a model-based approach. Both are integrated into NSGA-III, a multi-objective evolutionary algorithm, to explore Super-Resolution architectures optimizing three objectives: maximizing PSNR while minimizing parameter count and floating-point operations. We analyze each method’s impact on final architecture quality. SynFlow offers faster and more reliable estimations compared to XGBoost, which is slower and less accurate. However, XGBoost enables broader approximations of the Pareto front, making it suitable for more comprehensive trade-off exploration. Therefore, we recommend using SynFlow in scenarios where quick and resource-constrained searches are required, while XGBoost is more appropriate when the benefits of a wider trade-off analysis outweigh the additional computational cost.
Surrogate-assisted evolutionary algorithms (SAEAs) are a representative approach for expensive optimization problems, as solutions for expensive function evaluations (FEs) are prescreened or iteratively improved with the surrogate model of objective functions. These days, SAEAs are often designed to enhance convergence speed by utilizing previously evaluated solutions with good function values. Although this convergence-first design enables SAEAs to find relatively better solutions quickly at the beginning of the search, SAEAs tend to select quite similar solutions to each other for expensive FEs in the middle to the end of the search, resulting in the waste of expensive FEs. Accordingly, this work first incorporates a principle of tabu search into existing SAEAs as well as observes the behavior of the existing SAEAs and their extension. Tabu search extended for continuous optimization prevents to evaluation of similar solutions to previously evaluated ones. This promotes a variety of solution choices in SAEAs and realizes a sustainable performance improvement toward the end of the search. In the experiments held in this work, some SAEAs that have different search strategies and machine learning models for surrogate models are selected, and their extended versions incorporating tabu search are tested in a suite of single-objective continuous optimization benchmarks under an expensive scenario.
Pareto local optimal solution networks (PLOS-nets) model fitness landscapes as graphs where nodes are Pareto local optima and edges account for standard neighbourhood relationships. PLOS-nets capture the global structure of multi-objective combinatorial landscapes and provide a rich set of features. Published results so far focus on small problems. This article proposes new modelling and sampling methodologies to scale up PLOS-nets to larger problem sizes. Our sampling approach is based on the effective multi-objective random one-bit climber (moRBC). In our coarse-grained model, nodes are sets of Pareto local optimal solutions grouped according to the logarithm of their Pareto rank (following non-dominated sorting). For modelling edges, we consider the original neighbourhood edges, but also propose a new definition representing soft-restarts from Pareto local optima. We analyse and visualise our models on a set of combinatorial landscapes with tuneable ruggedness and number of objectives (MNK-landscapes). The models provide landscape features that correlate with the performance of multi-objective optimisation algorithms, and can be used in algorithm recommendation settings. The model with neighbour edges slightly outperforms the soft-restart model (in terms of predictive power), but its substantial computational overhead makes the soft-restart model a promising approximation.
