Deep Learning (DL) models, despite high accuracy, often degrade under domain shifts and image perturbations. This has frequently been observed in histopathology, where differences in slide preparation protocols may result in dramatically different tissue aspects, independently from any underlying biological phenomena. Some solutions, such as Lipschitz-constrained networks, offer formal robustness guarantees at the cost of added complexity. Others, such as foundation models, rely on general-purpose features learned on large datasets to provide better generalization. In this work, we explore genetic programming as an alternative, examining the inherent robustness of image classification programs evolved via Multi-Modal Adaptive Graph Evolution (MAGE). Using the PatchCamelyon histopathology dataset, we compare evolved programs against the Virchow2 foundation model, a standard ResNet-18 and a 1-Lipschitz ResNet-18 under Gaussian noise, Poisson noise, and brightness perturbations. We also study more realistic sources of noise such as Jpeg compression and staining alterations performed using AugmentHE. On unperturbed data, Virchow2 obtains the best performance by a large margin, but is subject to accuracy drops in most extreme noise regimes. MAGE programs, evolved without data augmentation or explicit robustness objectives, exhibit inherent resilience comparable to, and sometimes surpassing Virchow2 and 1-Lipschitz ResNet-18. Moreover, as the evolutionary search yields a diverse population, it enables the selection of highly robust programs suited to specific perturbations.
Genetic Programming (GP) has recently been applied to evolve aggregation functions in Federated Learning (FL), instead of fixing them in advance. FedGP represents aggregation rules as symbolic expressions that evolve during training and has shown competitive performance under non-IID data. Progress on GP-based aggregation, however, has been constrained by two practical barriers: limited reproducibility in FL experimentation and the absence of a public FedGP implementation. This paper presents FASTER, an open-source platform that provides the first public implementation of FedGP and integrates it with standard baselines (FedAvg, weighted FedAvg, FedProx, and FedNova). FASTER provides a browser- and CLI-based configuration-driven workflow for controlled FL experiments with built-in or custom datasets, IID and non-IID partitions, client scheduling, churn simulation, live metric tracking, and structured artifact export. Each run is stored with its configuration, logs, metrics, metadata, checkpoints, and, for FedGP, the individuals evolved across generations. In a demonstration on FashionMNIST with non-IID data and client churn, FedGP outperforms all fixed baselines with statistically significant gains; GPU batch evaluation reduces the previously reported roughly tenfold wall-clock overhead to about 3.3× relative to fixed-rule methods. By placing evolved and fixed aggregation rules under identical controlled conditions, FASTER turns FedGP into a reusable, testable, and extensible research object for the evolutionary computation community.
Ensuring fairness in machine learning remains a persistent challenge, as the most suitable model configurations can vary substantially under different fairness conditions. Our prior work provided a preliminary investigation into how jointly optimizing algorithm selection and hyperparameter tuning affects model fairness and accuracy. Building on that, this study further advances the research in four key aspects. First, we formally propose AutoFairGE, a fairness-aware AutoML framework based on Multi-Objective Grammatical Evolution, which automatically searches for machine learning models that balance accuracy and fairness. Second, we test AutoFairGE under a range of fairness conditions, considering multiple metrics for group and individual fairness in addition to different sensitive features. Third, we conduct a comprehensive empirical evaluation, comparing AutoFairGE with a broad set of baselines, including traditional fairness methods, modern fairness-aware AutoML methods, and a hyperparameter optimization method. Finally, we provide an in-depth analysis of the ML algorithms included as primitive functions in AutoFairGE and their influence on model accuracy and fairness. Our experiments demonstrate that AutoFairGE achieves competitive accuracy-fairness trade-offs while offering diverse solution sets adaptable to various fairness conditions.
In conventional fitness-based Genetic Programming (GP), parent selection is driven mainly by fitness and has limited access to relationships between programs beyond their observed performance. This work investigates whether pretrained code embeddings can provide complementary guidance for parent selection in symbolic regression. We introduce two mechanisms that combine fitness with distances computed in the latent spaces of CodeBERT and Code2Vec. Probabilistic Embedding-Guided Selection (PEGS) alternates between fitness-based and embedding-based selection under fixed or dynamically scheduled pressure. Two-Stage Embedding-Guided Selection (TSEGS) applies fitness selection followed by an embedding-based criterion aimed either at intensifying the search around elite solutions or at preserving diversity. The methods are evaluated on reduced SRBench datasets against standard GP and, in a targeted experiment, an NSGA-II baseline with constant optimization. Results show that embedding-derived information can improve held-out test performance, but its effectiveness depends on the embedding model, selection mechanism, and role assigned to distances. The diversity-oriented TSEGS criterion is most consistently beneficial with Code2Vec, whereas other configurations show more limited or problem-dependent effects. Model-size effects are mixed, and in the targeted NSGA-II experiment the gains are accompanied by larger median trees. A representative-run analysis further indicates that embedding-guided selection does not uniformly increase exploration. Overall, pretrained code embeddings provide a potentially useful but non-universal signal for GP selection, with benefits that must be weighed against configuration sensitivity, model-size effects, and substantial, implementation-dependent computational overhead.
This article reviews EvoGym, a lightweight Python framework for simulating voxel-based robots in two-dimensional environments. It describes the simulator's core features, including its Gymnasium-compatible API, diverse benchmark tasks, and flexible tools for designing custom environments. The review highlights EvoGym's accessibility, computational efficiency, and suitability for research and education in evolutionary robotics and reinforcement learning. Overall, EvoGym is a practical platform for research and education in embodied AI.
