ABSTRACT. This paper introduces a deep learning-based approach to the sonification of daily weather forecasts by mapping meteorological data to musical features via a latent emotion space grounded in psychological theory. Unlike traditional rule-based mappings, our model consisting of Variational Autoencoder and Feed-forward Neural Network learns associations between weather and music using numerical weather prediction data and emotion-annotated datasets. The system generates music that reflects weather conditions through features such as tempo, note duration, pitch range, and mode. Listening evaluations revealed that participants could distinguish broad weather categories (e.g., favourable vs. unfavourable), though finer distinctions proved challenging. This work highlights music’s potential as an intuitive and inclusive medium for communicating weather information, particularly for blind, visually impaired, and neurodivergent users, and supports engagement in STEAM (Science, Technology, Engineering, Arts, and Mathematics) education.

ABSTRACT. Sound hybridization is a technique that combines the characteristics of multiple sounds to generate new sonic textures, with applications ranging from scientific analysis to timbral synthesis and assisted musical composition. In this work, we propose a matching-pursuit-based approach for adaptive sound hybridization, which avoids a static segmentation of the target and enables a more flexible and coherent reconstruction. The method iteratively selects sound atoms from a reference database, optimizing the synthesis process in terms of both timbral and temporal-structural coherence. We demonstrate that our approach preserves the structure of the target while introducing controlled variations and analyze the impact of different feature spaces on the quality of the hybridization. The results show that the proposed strategy effectively balances fidelity to the target with an emphasis on the distinctive characteristics of the search space from which the sound objects involved in the hybridization process are selected. Our method prioritizes the aesthetic and creative potential of sound hybridization, offering a tool for artistic exploration.

ABSTRACT. Music loops are in increasing demand for music creation as well as live performance. There are numerous free and commercial options available for procuring and using loops from libraries and community repositories. A challenge commonly encountered by musicians is searching for loops that match their sonic requirements within the large space of available options. Musicians typically rely on metadata associated with loops, such as key and tempo, to narrow the search space and subsequently make selections using their own perceptual judgments. This process can be time-consuming and effort-intensive. In our work, we investigate whether AI audio models can assist musicians in loop search tasks. We develop LoopMatcher, a workflow that integrates two complementary AI models - one for efficient embedding-based retrieval (VGGish) and the other for perceptual similarity refinement (CDPAM) - to automatically narrow the search space and rank candidate loops given a reference loop. In a proof-of-concept validation study, we observe a Spearman correlation of 0.68 between algorithmic rankings and human perceptual judgments. Additionally, randomly selected loops consistently rank lowest, judged by both the system and participants. The results of our work provide encouraging outcomes that indicate strong correlations between our approach and human perceptual judgments, demonstrating that AI-assisted loop search shows promise and merits further investigation.

ABSTRACT. Recent advances in quantum computing have opened new possibilities for modeling high-dimensional and entangled structures in musical data. This paper introduced a Quantum Variational Autoencoder (QVAE) architecture for symbolic music generation, leveraging quantum latent spaces to capture superposed and correlated melodic patterns beyond the expressive capacity of classical generative models. By encoding MIDI-based musical sequences into parameterized quantum circuits, the proposed framework enabled the exploration of superposed latent states, facilitating multiple potential melodic continuations within a unified probabilistic representation. The study evaluated QVAE against conventional Variational Autoencoders using both objective metrics, including latent diversity, reconstruction accuracy, and sequence entropy, and subjective listening tests assessing musical coherence and creativity. Experimental results demonstrated that QVAE achieved enhanced diversity in generated melodies while maintaining stylistic consistency, suggesting that quantum-inspired latent modeling can serve as a novel pathway for computational creativity in music information retrieval. Beyond generation, the work discussed the implications of quantum latent structures for music representation learning, proposing a foundation for future hybrid quantum-classical MIR frameworks capable of handling musical complexity through the principles of superposition and entanglement.

ABSTRACT. A song narrates a story through its lyrics, rhythm, and structure. Understanding the song structure is beneficial to provide valuable insights of the song’s anatomy. Therefore, song structural segmentation, one of the main tasks of music information retrieval (MIR), has drawn attention and has been popular in the music industry with relevant and diverse applications of song cutting, song recommendation, song classification, etc. However, automatically detecting the structure of song is challenging as different music genres present different principles for song structure without a general rule. In this paper, we propose a deep learning model based on an encoder–decoder architecture, combined with hand-crafted acoustic features, for end-to-end song structure segmentation. We train and evaluate our proposed model on 800-song self-collected dataset and a benchmark dataset of the Beatles which present a wide range of music genres. We also construct two baseline systems which use the conventional Laplacian segmentation (LS) method and a combination between Laplacian segmentation and multilayer perceptron network (LS-MLP) to compare with the proposed encoder-decoder network. Experimental results on the datasets demonstrate that our proposed encoder-decoder network outperforms two baselines, especially for detecting the boundaries of audio segments. Furthermore, our model proves effective to detect the label for each segment. The high performance evaluated on diverse music genres shows potential to apply the proposed model for music structuring applications.

ABSTRACT. Western music theory concepts underlie the compositional processes of much of the music consumed today and consequently serve as the conceptual underpinnings of much of the musical data used to train music generative models. Applying the interpretability technique of probing, Wei et al. studied how well basic musical concepts such as scales, time signatures, and simple triadic harmonic progressions are encoded in the latent spaces of the Jukebox and MusicGen generative models. Through constructing datasets isolating these concepts, the researchers trained simple classifiers to classify elements of these concepts based on the latent representations. We extend their work by creating datasets that focus on five higher-level and more advanced Western music theory concepts: polyrhythms, dynamics, seventh chords, mode mixture, and secondary dominants. Focusing our experiments on the same Jukebox and MusicGen models, we train probing classifiers on the datasets' latent representations. We find that these advanced concepts are encoded in their latent spaces and that secondary dominant classification is a more difficult probing task than our other music theory tasks.

ABSTRACT. Music Emotion Recognition (MER) is a challenging task considering the nuances of defining emotions. While unimodal models provide a good baseline for MER, multimodal models are becoming fundamental to provide an in-depth description of emotions. Leveraging on the multimodal MERGE dataset, we investigate the power of audio-related deep embeddings, lyrics informed features, and music-aware cues in providing an informative set of features for low-impact computational learning models. Results confirm that multimodal fusion outperforms unimodal approaches. Moreover, different experiments highlighted the positive contribution of genre metadata and the potential use of harmonic features for real-time computationally low-impact applications. These findings confirm the importance of multimodal integration for robust and interpretable emotion recognition systems, while opening up future directions, including advanced feature fusion, user-specific model adaptation (user-tuning), and multi-label emotion representation.

ABSTRACT. Automated selection of the best-performing algorithm for a given optimisation problem is crucial. However, how effectively automated algorithm selectors perform on problem classes beyond their training set remains an open area of investigation. While prior research has examined the impact of different problem representations and benchmarking suites on the ability of selectors to perform well on unseen problem classes, the influence of optimisation algorithm portfolio composition and machine learning model choice remains underexplored. This study investigates how these factors influence cross-benchmark performance of selectors in single-objective black-box optimisation. Problem instances are generated from two distinct benchmarking suites, i.e., the Black-Box Optimisation Benchmarking and Sinha-Malo-Deb, and represented using Exploratory Landscape Analysis features. Ninety-nine portfolios, each containing three to seven optimisation algorithms, and six machine learning models are evaluated under a Leave-One-Suite-Out evaluation scenario. The results show that portfolio composition strongly influences both Oracle-Baseline performance gap (the difference between the performance of an ideal selector and a baseline selector) and portfolio complementarity (how well algorithms within a portfolio complement each other). These factors, in turn, govern the ability of selectors to generalise beyond their training problem set. Furthermore, the choice of machine learning model can substantially affect how much of Oracle-Baseline performance gap is closed. Moreover, some machine learning models leverage portfolio complementarity more effectively to perform well across benchmarking suites. These findings highlight that careful selection of optimisation algorithms within a portfolio, combined with appropriate machine learning models, is key to robust automated algorithm selection.

ABSTRACT. In recent years, Hebbian Learning (HL) was employed in several Reinforcement Learning (RL) tasks to maintain high plasticity within the models, while Quality-Diversity (QD) algorithms have been exploited to retrieve diverse high-performing solutions. In this work, we propose a combination of QD algorithms with the recently introduced Neuron-Centric Hebbian Learning (NcHL) to tackle RL tasks. QD methods aim to evolve diverse, high-performing behaviors, offering enhanced robustness and potentially improved interpretability, compared to conventional optimization. Moreover, NcHL introduces a scalable HL approach based on neuron-centric local plasticity rules, enabling on-device adaptation without gradient-based updates. We evaluate the proposed framework on three standard RL benchmarks, CartPole, MountainCar, and LunarLander, using three state-of-the-art QD algorithms: MAP-Elites (ME), CMA-ME, and CMA-MAE. Experimental results demonstrate that combining QD with NcHL facilitates the emergence of heterogeneous yet effective control strategies. Behavioral analyses further highlight diverse plasticity dynamics and task-specific adaptation.

ABSTRACT. Many embedded applications have strict energy, memory, and time constraints, making neural network inference particularly challenging. Recently, a novel NN architecture called Fast Feedforward Networks (FFFs) has been proposed to achieve inference with extremely lightweight computational demands and minimal latency. However, FFFs still face two key limitations: (1) their memory footprint remains relatively high despite fast inference, and (2) their architectures are restricted to balanced forms, where each branch has the same depth and each leaf has the same width. This can limit their scalability, performance, and potential for further improvements in inference speed. In this work, we address these challenges by introducing a novel Neural Architecture Search pipeline that explores imbalanced FFF architectures using multi-objective evolutionary optimization. Specifically, we utilize NSGA-II, setting the objectives as accuracy and total number of model parameters, along with Grammatical Evolution, a method particularly effective for tree-structured models. We conduct comprehensive experiments on three different IoT datasets, benchmarking our method against standard (i.e., balanced) FFFs and feedforward models in terms of accuracy, Multiply-Accumulate Operations, and parameter count. Furthermore, we analyze the Pareto-optimal architectures found by NAS, to gain insights into their structural properties. Our experiments demonstrate that the proposed approach can achieve an up to 13% accuracy gain compared to FFFs with the same model size, and reduce model size by up to 10x compared to FFFs with equivalent performance, all while achieving lower or comparable inference cost.

ABSTRACT. Large Language Models remain vulnerable to adversarial prompts that elicit toxic content even after safety alignment. We present a black-box evolutionary framework that tests model safety by evolving prompts in a synchronous, steady-state (mu+ lambda) loop. The system employs a diverse operator suite, including lexical substitutions, negation, back-translation, paraphrasing, and two semantic crossover operators, while a moderation oracle provides fitness guidance. Under a fixed generation budget, a few-shot global rewrite operator achieves the highest progress per evaluated prompt but plateaus at substantially lower best-of-run toxicity than our engineered lexical operators, which more reliably push populations toward high-toxicity regimes. Operator-level analysis reveals significant heterogeneity, as lexical substitutions offer the best yield–variance trade-off, semantic-similarity crossover acts as a precise low-throughput inserter, and global rewrites exhibit high variance with elevated refusal costs. Using elite prompts evolved on LLaMA 3.1 8B, we observe practically meaningful but attenuated cross-model transfer. Toxicity drops by roughly half on most targets, with smaller LLaMA 3.2 variants showing the strongest resistance and some cross-architecture models (e.g., Qwen and Mistral) retaining higher toxicity. Overall, our results indicate that small, controllable perturbations serve as reliable vehicles for systematic red-teaming, while defenses should anticipate cross-model prompt reuse rather than focusing solely on single-model hardening.

ABSTRACT. Test-time search is an attractive alternative to parameter scaling for improving the reasoning ability of large language models (LLMs). We introduce Evo-Reasoner, an evolutionary framework that optimizes the content of an LLM’s chain-of-thought rather than its prompts or weights. For each problem, an AUTHOR LLM generates structured, tagged solutions, which are evolved under a MAP-Elites quality-diversity archive using cost-aware Elo ratings assigned by a separate JUDGE LLM. Specialized mutation and crossover operators act on reasoning fragments, and a Global End-of-Generation Judge consolidates convergent evidence into a final answer. On GSM8K-style math word problems and a 100-question subset of MMLU, Evo-Reasoner yields large gains for 7B-class models: DeepSeek-7B improves from 14% to 58% accuracy on GSM8K, and Qwen-7B rises from 52% to 93%, surpassing a 20B open-source baseline; both backbones also improve substantially on MMLU. Evolution further produces shorter, more coherent, and better-verified reasoning traces. These results show that structured evolutionary refinement at inference time can substantially amplify the reasoning capabilities of small LLMs without parameter updates or additional training data.

ABSTRACT. In this work, we present an automated framework for discrete interpretable prompt optimization by combining Evolutionary Algorithms (EAs) with Large Language Models (LLMs). Building on the recently proposed EvoPrompt methodology, we adopt a Genetic Algorithm (GA) and Differential Evolution (DE) to optimize prompts for vision-language models, specifically focusing on CLIP prompt templates. Our approach leverages the Alpaca LLM to perform evolutionary operations, maintaining prompt coherence while exploring diverse prompt formulations. We evaluate the effectiveness of our framework across three datasets: ImageNet-1k, 102Flowers, and FGVC-Aircraft-2013b, using a similarity-based fitness function. Our experimental results demonstrate that larger population sizes and specific selection strategies enhance prompt optimization. This work highlights the potential of combining evolutionary search with LLMs to improve automated prompt engineering while preserving interpretability.

ABSTRACT. Benchmarking in continuous black-box optimisation is hindered by the limited structural diversity of existing test suites such as BBOB. We explore whether large language models embedded in an evolutionary loop can be used to design optimisation problems with clearly defined high-level landscape characteristics. Using the LLaMEA framework, we guide an LLM to generate problem code from natural-language descriptions of target properties, including multimodality, separability, basin-size homogeneity, search-space homogeneity and global–local optima contrast. Inside the loop we score candidates through ELA-based property predictors. We introduce an ELA-space fitness-sharing mechanism that increases population diversity and steers the generator away from redundant landscapes. A complementary basin-of-attraction analysis, statistical testing and visual inspection, verifies that many of the generated functions indeed exhibit the intended structural traits. In addition, a t-SNE embedding shows that they expand the BBOB instance space rather than forming an unrelated cluster. The resulting library provides a broad, interpretable, and reproducible set of benchmark problems for landscape analysis and downstream tasks such as automated algorithm selection.

ABSTRACT. This paper suggests a problem that might be impossible to solve with evolutionary algorithms despite its easy formulation. “Sab- rina’s Game” consists of a matrix with numbers, and each number needs to be covered by an oval together with a neighbouring number. The ob- jective value of an oval is the largest number it covers minus the smallest number it covers, and the total summed ovals’ objective values is the solution’s objective value. This value is to be minimized, but a solution is only valid if every number in the matrix is covered by an oval. These requirements might make the problem both nontrivial and unamenable for evolutionary algorithms, due to unsampleability of initial random solutions

ABSTRACT. The HP protein folding problem has no known determinis- tic polynomial time sampling algorithm, and mutations known for the problem are often problematic because they can produce invalid confor- mations.In this first exploration, I will conjecture that valid conforma- tions could be fractally distributed in the search space. This distribution might explain why we encounter the problems with random sampling and mutation as we do in this NP-complete problem

ABSTRACT. Designing conformal cooling channels (CCC) for injection molding is challeng-ing because multiple geometric, thermal, and mechanical constraints interact, and improving one performance measure often degrades another. While multi-objective evolutionary optimization can generate strong trade-off designs, indus-trial selection is often based on manual inspection, which limits transparency and the reuse of decisions. This paper investigates how interactive knowledge discov-ery can support post-optimization decision-making. We solve four two-objective CCC design problems using NSGA-II, employing a high-fidelity Moldex-3D–based ANN surrogate trained on 160 feasible designs to enable fast evaluation. The resulting Pareto-approximate sets are analyzed with MIMER, which links visual exploration and pattern mining to connect preferred objective-space regions to interpretable rules and influential design variables. Compared with an inde-pendent expert assessment, MIMER consistently identifies the same key varia-bles—especially channel diameters, distance to the cavity surface, and gate loca-tion—and produces insights aligned with expert reasoning. It also reveals stable cross-run relationships not obvious from scatter plots, such as gate-dependent placement corridors and diameter ranges associated with efficient, uniform cool-ing. Overall, the results indicate that interactive knowledge discovery can serve as a semi-automatic analyst for CCC optimization, improving transparency, repro-ducibility, and knowledge reuse in injection molding.

ABSTRACT. Optimizing Mixture-of-Experts (MoE) communication topology presents a vast combinatorial search space, seemingly ideal for evolutionary algorithms (EAs). We conduct compute-matched experiments comparing a genetic algorithm against random search across six independent runs. Surprisingly, random search matches or exceeds EA performance, achieving lower perplexity in four of six paired comparisons. This result holds across tested mutation configurations and multiple algorithm designs. We trace the cause to a weakly informative fitness landscape: population analysis reveals structural diversity collapses while fitness variance remains unchanged.

ABSTRACT. Machine learning models trained on biased data can perpetuate discrimination based on sensitive attributes such as race or gender. While numerous bias mitigation techniques exist for tabular data, most address only binary classification with a single binary sensitive attribute. This work proposes FairGenes, an evolutionary algorithm that constructs sequential unbiasing pipelines capable of handling non-binary sensitive attributes by leveraging existing binary-oriented techniques. FairGenes encodes sequences of data transformations as variable-length genotypes and employs surrogate classifiers to optimise fairness independently from the target classifier. The one-vs-all decomposition strategy and lexicographic selection enable the algorithm to balance multiple fairness and performance objectives across multi-class protected features.

ABSTRACT. Diabetic Foot Ulcers (DFU) require timely severity assessment to guide treatment and follow-up. This paper investigates automated four-class DFU severity classification (Wagner-Meggitt Grades 1-4) using transfer learning on a public Kaggle dataset. The dataset includes 10062 images with roughly 446-450 subjects per grade. To ensure leakage-free evaluation, we extract subject IDs from standardized filenames and employ subject-wise 5-fold cross-validation. We compare five ImageNet-pretrained CNNs: ResNet18, ResNet50, DenseNet121, GoogLeNet, and MobileNetV2. We further study the effect of dataset reliability by evaluating Kaggle as provided and ( an expert-verified, quality-screened (QC) subset obtained by removing mislabeled and low-quality samples (1539 images). Augmentation is applied only during training with two strengths (4 vs. 10 augmented copies per image), while validation and test sets remain unchanged. DenseNet121 attains 62.78% accuracy on the Kaggle setting, whereas QC with stronger augmentation achieves 89.23% accuracy, indicating that expert verification and controlled augmentation substantially improve robustness for DFU severity classification.

ABSTRACT. Text-only training is a promising new machine learning paradigm for training multimodal models without requiring data from every modality. However, despite the potential of text-only training, to date few studies have explored its use as an approximation of missing data for supervised learning in data-scare environments. In this work we define the kinds of problems suited for text-only training and examine techniques to acquire or design text-based training data. We address the modality gap's role in the performance of text-only training, and present a case study on classifying subjective audio timbre descriptions based on three kinds of text-only training data and six augmentation methods on eight audio-timbre datasets. Our work shows that the text-only training paradigm successfully trains audio classifiers without audio and opens the door to future work in examining the effectiveness of text-only training for supervised machine learning problems without available datasets.

ABSTRACT. This article addresses how “meaning” and “creativity” ought to be conceptualized within a multi-agent ecology at a time when AI is being rapidly embedded in the creative industries, shifting music pro- duction from performance-centered practices to design/prompt-driven workflows. The principal gap in the literature is the overemphasis on the singular genius/subject and a human-centered narrative, along with the intrinsic qualities of the “work,” while recommendation algorithms, prove- nance/traceability, and institutional selection rules remain insufficiently incorporated into prevailing models; as a result, the debate is often locked in the “are machines creative?” binary. The claim is as follows: AI is a powerful accelerator/amplifier; however, creative value and musical mean- ing are relational properties that are embodied, context-sensitive, and emergent through social legitimation within composer––––feedback loops of distributed intentionality. Accordingly, “creativity” is not an intrinsic attribute but a processual capacity constructed through commitments anchored by provenance, embodied/enactive interaction, and community- based acceptance in circulation; “meaning” is an ecological emergence that materializes through articulation to the output’s production/distribution traces, usage practices, and cultural ontologies. Method/scope: Through a theoretical–synthesis that mobilizes motifs from complexity science (emergence, feedback, multiscalarity), the production–––chain is modeled jointly at the micro (performance, micro-timing) and macro (circulation, content policies, provenance) layers. Findings: (i) AI accelerates produc- tion and expands the expressive space; (ii) the “work” is not a fixed essence but a process node within a traceable chain of operations; (iii) “costly commitments” (live-performance risk, craft labor, open data/code, trace- able production) generate signals of trust and differentiation; (iv) success depends as much on early exposure and network position as on intrinsic qualities; (v) artificiality/provenance framings shape listener preferences; (vi) the meaningfulness of machine outputs is conditioned by provenance transparency, embedding within interpretive communities, and indicators of embodied/enactive coordination. Contribution: The article advances, theoretically, a “work = process” ontology, a distributed architecture of creativity, and a scenario-oriented/probabilistic epistemology of the ecological origins of meaning; methodologically, a dual-lens and mixed- methods standard that jointly addresses production (micro) and circula- tion/policy (macro) layers; and, in practice, distribution strategies that ethically manage the exploration–balance alongside a provenance-based trust architecture. Conclusion: The field should abandon the “creativity = essence” assumption in favor of an ecology-and-network framework; cur- ricula and policy should center hybrid (human + AI) production literacy, competence in algorithmic circulation, and provenance ethics; and future research should proceed with culturally adapted modeling and evaluation settings and with “hybrid bundles” (work analysis + circulation metrics + embodied/enactive indicators).

ABSTRACT. This study explores the capacity of generative artificial intelligence (Gen AI) to contribute to the construction of peace narratives and the revitalization of musical heritage in Mali. The study has been made in a political and social context where inter-community tensions and social fractures motivate a search for new symbolic frameworks for reconciliation. The study empirically explores three questions: (1) how Gen AI can be used as a tool for musical creation rooted in national languages and traditions; (2) to what extent Gen AI systems enable a balanced hybridization between technological innovation and cultural authenticity; and (3) how AI-assisted musical co-creation can strengthen social cohesion and cultural sovereignty. The experimental results confirm that Gen AI, embedded in a culturally conscious participatory framework, can act as a catalyst for symbolic diplomacy, amplifying local voices instead of standardizing them. However, challenges persist regarding the availability of linguistic corpora, algorithmic censorship, and the ethics of generating compositions derived from copyrighted sources.

ABSTRACT. We present an evolutionary approach to optimize and generate images by evolving prompt embeddings, the tensor representations produced by the text encoder of a diffusion model. Instead of editing text prompts, the search operates directly in the continuous embedding space, while the diffusion generator remains fixed. Optimization objectives are defined through CLIP Image Quality Assessment (CLIP-IQA), covering both native and custom image scoring criteria such as "naturalness", "complexity", or "happiness". Using the SDXL-Turbo model, we evaluate setups with three, four, five, and ten objectives. A single-objective Genetic Algorithm (GA) maximizing the summed scores serves as a baseline, compared against our implementation of the Non-dominated Sorting Genetic Algorithm III (NSGA-III) for multi- and many-objective optimization. Results demonstrate that embedding-space evolution is feasible and effective: NSGA-III consistently outperforms the GA in both quality and diversity, revealing interpretable trade-offs between visual criteria. All code, data, and evaluation tools are released as part of the open-source Evolutionary Diffusion Framework.

ABSTRACT. Understanding how language models (LMs) acquire and refine semantic meaning during training is an open challenge in natural language processing (NLP). Prior work has largely focused on static evaluations of pretrained models, providing limited insight into the temporal dynamics of meaning representation. In this study, we conduct the most fine-grained longitudinal analysis to date of semantic similarity tracking in autoregressive language models. Using models from the Pythia family (70M and 410M parameters), we systematically evaluate six complementary metrics— AUPRC, AUROC, Kendall’s Tau, KT-Cosine, KTKendall, and Spearman correlation—across three established human-judgment datasets (SimLex-999, WordSim-353, MTurk-771). Our results reveal consistent patterns: (i) early training stages exhibit sharp improvements in alignment with human similarity judgments, (ii) larger models (410M) achieve consistently higher correlations than smaller models (70M), and (iii) certain metrics (e.g., KT-KT) peak early before gradually declining, indicating non-monotonic dynamics of semantic structure formation. We release a largescale dataset of over 160,000 evaluation traces (JSON, CSV, and plots) covering 80,000 training steps, enabling reproducibility and further exploration of semantic dynamics. To our knowledge, this is the first dataset capturing step-by-step semantic evolution across model scales. These findings not only extend prior work on meaning dynamics in LLMs but also provide a valuable benchmark resource for the community

ABSTRACT. For decades, metaheuristics have been dominated by natural system analogies, leaving cognitive science frameworks largely untapped. In this paper, we introduce the Reinforcement Learning Agent (RLA) algorithm, which reconceptualizes global optimization as a single agent navigating a search space via a Markov Decision Process (MDP). RLA employs Q-learning to adaptively select from seven search operators, including Lévy flights, chaotic maps, opposition-based learning, and adaptive differential evolution, guided by a layered memory system. Extensive evaluation on the CEC 2020 benchmark suite (covering dimensions from 5 to 20) demonstrates RLA’s strong performance, yielding an overall Friedman rank of 1.585 and surpassing its closest competitor, Adaptive Guided Differential Evolution with Self-adaptive Knowledge (AGSK, 2.716). RLA finds global optima for functions f1−f7 and remains competitive across all functions, surpassing L-SHADE (Linear Population Size Reduction Success-History Based Adaptive Differential Evolution) and its variants for dimensions ≥ 10 (Wilcoxon p < 0.05). With high computational efficiency and excellent scalability, requiring fewer evaluations than population-based methods, RLA emerges as a powerful tool for complex optimization. Its cognitive-inspired design opens new avenues for learning-driven metaheuristics.

ABSTRACT. Preference learning amounts to ranking items based on user preferences or their relevance. It has been demonstrated that preference learning with subjective preferences is well-suited to machine learning using neuroevolution rather than more traditional stochastic gradient descent methods. Specifically, neuroevolution as implemented in NEAT is used as a basis in the RankNEAT algorithm, which performs preference learning by solving the ranking problem for subjective labels. Although RankNEAT shows promise, the algorithm is limited by the relatively slow execution speed of its CPU-based implementation of NEAT. To handle NEAT's slow execution speed, the TensorNEAT algorithm was developed. TensorNEAT tensorizes the workload of NEAT and leverages GPU acceleration to speed up NEAT; however, TensorNEAT does not perform ranking.

To handle this challenge, we develop the TensorRankNEAT algorithm. TensorRankNEAT performs preference learning (like RankNEAT) using tensors that enable fast GPU implementations (like TensorNEAT). TensorRankNEAT thus adresses the computational challenges of using NEAT for preference learning. In this work, this new algorithm is developed and tested. In experiments with the AGAIN dataset, we find that the performance of TensorRankNEAT compares favorable to previous experimental results for RankNEAT found in the literature. At the same time, due to its tensorization and GPU implementation, it is substantially faster, showing its potential both for other applications in preference learning as well as further progress in preference learning by means of neuroevolution.

ABSTRACT. A multi-objective, hardware-aware Neural Architecture Search (NAS) framework is introduced and evaluated for efficient design of convolutional neural networks (CNN) on the Hailo-8L accelerator. The search for CNNs that exhibit good accuracy-latency trade-offs on image classification tasks is conducted using NSGA-II and a single-objective genetic algorithm that imposes a latency constraint. To reduce execution time, we adopt a pretrained supernet, introduce an optimized evolution-training interaction, and develop a latency predictor based on a surrogate model. The framework is primarily intended to construct CNN architectures optimized for Hailo-8L and user-provided application-specific datasets. Nevertheless, to facilitate a direct comparison with existing methods, we evaluated it on standard image classification benchmarks, specifically CIFAR-10, CIFAR-100, and ImageNet-100.

ABSTRACT. Training semantic segmentation networks on synthetic data enables automatic annotation but suffers from domain gaps when deployed on real images. Data augmentation and style transfer can bridge this gap, yet determining the right data transformations and their sequence is difficult. We formulate the selection of style transfer pipelines as a modified traveling salesman problem where a terminal node ends the sequence. Using multi-objective genetic algorithms, we optimize pipelines to balance structural coherence and style similarity to target domains. The main challenge lies in evaluation: distributional metrics require generating many samples, making them expensive for optimization. Instead, we use paired-image metrics on individual samples during evolution. After optimization, we evaluate the resulting Pareto front using distributional metrics and segmentation performance. We apply this approach to synthetic-to-real domain adaptation from GTA5 to Cityscapes and ACDC datasets under adverse conditions. Results demonstrate that evolutionary algorithms can propose various augmentation pipelines adapted to different objectives. The contribution of this work is the formulation of style transfer as a sequencing problem suitable for evolutionary optimization and the study of efficient metrics that enable feasible search in this space. The source code is available at: https://anonymous.4open.science/r/MOOSS-FE26.

ABSTRACT. This study investigates the effect of alternating growth and pruning phases on the evolution of neural architectures using the Evolutionary eXploration of Augmenting Memory Models (EXAMM). Our objective was to determine if this structured evolutionary process could guide the search towards more compact models without sacrificing predictive performance. We compared EXAMM-evolved architectures having traditional and five modern memory cells (simple, UGRNN, MGU, GRU, Delta-RNN, and LSTM) on two distinct time-series datasets: aviation flight recorder data and coal-fired power plant data. The alternating phases were controlled by enabling growth-promoting or growth-reducing mutation operations based on trigger frequencies of 50, 100, or 200 generated genomes, resulting in a 3×3 set of nine unique grow-shrink configurations. Results show that while there was no statistical difference in terms of model performance to a baseline without phases, the experiments with higher pruning to growth ratios had statistically significant decreases in model size, making this an effective method for controlling network complexity.

ABSTRACT. Evolving neural network architectures is a computationally demanding process. Traditional methods often require an extensive search through large architectural spaces and offer limited understanding of how structural modifications influence model behavior. This paper introduces NeuroEVOlution through Geometric Semantic perturbation and Population based Training (NEVO-GSPT), a novel Neuroevolution algorithm based on two key innovations. First, we adapt geometric semantic operators (GSOs) from genetic programming to neural network evolution, ensuring that architectural changes produce predictable effects on network semantics within a unimodal error surface. Second, we introduce a novel operator (DGSM) that enables controlled reduction of network size, while maintaining the semantic properties of GSOs. Unlike traditional approaches, NEVO-GSPT’s efficient evaluation mechanism, which only requires computing the semantics of newly added components, allows for efficient population-based training, resulting in a comprehensive exploration of the search space at a fraction of the computational cost. Experimental results on four regression benchmarks show that NEVO-GSPT consistently evolves compact neural networks that achieve performance comparable to or better than established methods in the literature, such as standard neural networks, SLIM-GSGP, TensorNEAT, and SLM.

ABSTRACT. Algorithmic discovery has traditionally relied on human ingenuity and extensive experimentation. Here we investigate whether a prominent scientific computing algorithm, the Kalman Filter, can be discovered through an automated, data-driven, evolutionary process that relies on Cartesian Genetic Programming (CGP) and Large Language Models (LLM). We evaluate the contributions of both modalities (CGP and LLM) in discovering the Kalman filter under varying conditions. Our results demonstrate that our framework of CGP and LLM-assisted evolution converges to near-optimal solutions when Kalman optimality assumptions hold. When these assumptions are violated, our framework evolves interpretable alternatives that outperform the Kalman filter. These results demonstrate that combining evolutionary algorithms and generative models for interpretable, data-driven synthesis of simple computational modules is a potent approach for algorithmic discovery in scientific computing.

ABSTRACT. Although Differential Evolution (DE) remains a fundamental optimization technique, its mutation strategies often strive to strike a balance between exploration and exploitation in high-dimensional, multimodal environments. Employing the paradigm of LLM-assisted optimization algorithm generation, this paper presents a novel mutation that is synthesized using a methodology that exploits Large Language Models (LLMs) for automated strategy design. An LLM is provided with a guiding prompt to produce these mutation heuristics. The stages of this prompt’s structure take into account historical DE variants, problem characteristics, and evolutionary principles, and the resulting mutation and Differential Evolution algorithm are extensively benchmarked on the Black-Box Optimization (BBO) suite for 10-, 30-, and 40-dimensional spaces. When incorporated into DE, the LLM-generated mutation continuously outperformed state-of-the-art variants (which included DE/rand/1 & 2, DE/best/1 & 2, DE/current-to-rand/1, JADE, and Union differential evolution mutation) across 30 independent runs per function. It emerged first in all tested configurations (10D, 30D, and 40D) and earned the best average ranking. Statistically significant differences between the analyzed algorithms are indicated by the obtained p-values (3.96E−18, 5.76E−19,and 2.36E−20), which demonstrates the superiority of the outlined strategy and reinforces it’s effectiveness and dependability. Together with DE, this new mutation operator unlocks a window to AI-enhanced evolutionary algorithm engineering, which promises scalable advancements in Black-Box optimization tasks.

ABSTRACT. Heuristic functions are essential to the performance of tree search algorithms such as A*, where their accuracy and efficiency directly impact search outcomes. Traditionally, such heuristics are handcrafted, requiring significant expertise. Recent advances in large language models (LLMs) and evolutionary frameworks have opened the door to automating heuristic design. In this paper, we extend the Evolution of Heuristics (EoH) framework to investigate the automated generation of guiding heuristics for A* search. We introduce a novel domain-agnostic prompt augmentation strategy that includes the A* code into the prompt to leverage in-context learning, named Algorithmic - Contextual EoH (A-CEoH). To evaluate the effectiveness of A-CeoH, we study two problem domains: the Unit-Load Pre-Marshalling Problem (UPMP), a niche problem from warehouse logistics, and the classical sliding puzzle problem (SPP). Our computational experiments show that A-CEoH can significantly improve the quality of the generated heuristics and even outperform expert-designed heuristics. The code is available: https://anonymous.4open.science/r/a-ceoh-evolution-of-heuristics-92FE

ABSTRACT. Interpretability is increasingly crucial in artificial intelligence, especially in domains like medicine where genetic programming, particularly grammatical evolution (GE), can produce potentially interpretable models. However, optimizing for predictive accuracy often leads to complex expressions that reduce interpretability.We propose a large language model (LLM)-guided local search mechanism, implemented as a mutation operator within GE, to enhance model interpretability in symbolic regression for medical prediction. The operator leverages the semantic reasoning and domain knowledge encoded in LLMs to refine candidate expressions through a parallel search mechanism complementary to traditional fitness-driven evolution. The experiments, conducted on a clinical dataset for glycated hemoglobin prediction, showed that three LLM-guided variants improved interpretability compared to baseline GE, although two showed signs of possible overfitting while the other generalized better. LLM-generated individuals demonstrated evolutionary viability throughout evolution and direct contribution to the final solutions. Comparison of proprietary and local LLMs revealed challenges in instruction adherence and creative exploration for smaller models. These results indicate that the proposed approach is a promising strategy for obtaining simpler, more interpretable, and more plausible models. However, this work is an exploratory proof-of-concept with limitations. Future work will focus on evaluating broader datasets, refining design choices and validating clinical relevance.

ABSTRACT. The emerging field of AI-generated art has witnessed the rise of prompt marketplaces, where creators can purchase, sell, or share prompts to generate unique artworks. These marketplaces often assert ownership over prompts, claiming them as intellectual property. This paper investigates whether concealed prompts sold on prompt marketplaces can be considered as bona fide intellectual property, given that humans and AI tools may be able to sufficiently infer the prompts based on publicly advertised sample images accompanying each prompt on sale. Specifically, our study aims to assess (i) how accurately humans can infer the original prompt solely by examining an AI-generated image, with the goal of generating images similar to the original image, and (ii) the possibility of improving upon individual human and AI prompt inferences by crafting human-AI combined prompts with the help of a large language model (LLM). Although previous research has explored AI-driven prompt inference and protection strategies, our work is the first to incorporate a human subject study and examine collaborative human-AI prompt inference in depth. Our findings indicate that while prompts inferred by humans and prompts inferred through a combined human-AI effort can generate images with a moderate level of similarity, they are not as successful as using the original prompt. Moreover, combining human- and AI-inferred prompts using our suggested merging techniques did not improve performance over purely human-inferred prompts.

ABSTRACT. This paper examines how contemporary South Asian digital artists perceive, resist, and re-imagine the role of generative artificial intelligence (AI) in creative practice. Drawing from thirty-four qualitative interviews conducted for the South Asian Digital Arts Archive (SADA, 2025), the analysis maps emerging conversations on authorship, ethics, posthumanism, and infrastructural inequality. Artists across South Asia reveal heterogeneous positions: some embrace AI as a collaborator, while others reject it as ecologically or epistemically violent. Through a thematic reading grounded in postcolonial digital-humanities theory, the paper argues that South Asian responses to AI are shaped less by technophilia or fear than by a decolonial impulse to localize data, preserve embodied knowledge, and foreground human-centered creativity.

ABSTRACT. This study explores artificial visual creativity, focusing on ChatGPT’s ability to generate new images intentionally pastiching original artworks such as paintings, drawings, sculptures and installations. The process involved twelve artists from Romania, Bulgaria, France, Austria, and the United Kingdom, each invited to contribute with three of their artworks and to grade and comment on the AI-generated versions. The analysis combines human evaluation with computational methods aimed at detecting visual and stylistic similarities or divergences between the original works and their AI-produced renditions. The results point to a significant gap between color and texture-based similarity and compositional, conceptual, and perceptual one. Consequently, we advocate for the use of a "style transfer dashboard" of complementary metrics to evaluate the similarity between pastiches and originals, rather than using a single style metric. The artists' comments revealed limitations of ChatGPT's pastiches after contemporary artworks, which were perceived by the authors of the originals as lacking dimensionality, context, and intentional sense, and seeming more of a paraphrase or an approximate quotation rather than as a valuable, emotion-evoking artwork.

ABSTRACT. This paper presents a novel approach to generative art integrating quantum computing principles with artificial life paradigms. We propose a self-organizing computational ecosystem where autonomous digital entities evolve, interact, and co-create through quantum-enhanced behaviors. By leveraging quantum phenomena such as superposition and entanglement, our framework generates emergent artistic expressions with enhanced complexity and natural unpredictability. We introduce the Quantum Particulate Automata as an implementation framework, demonstrating how quantum principles reshape generative art methodologies beyond classical computational limitations.

ABSTRACT. Interactive Evolutionary Computation (IEC) has long served as a powerful methodology for human-guided design optimization, enabling the iterative refinement of artifacts through user-in-the-loop selection. However, the reliance on human input for fitness evaluation poses significant limitations in scalability, consistency, and exploration depth. In this work, we present a fully autonomous, LLM-driven agentic AI system for evolutionary design — one that eliminates the need for human evaluators by integrating the generative and evaluative capabilities of Large Language Models (LLMs) and Visual Large Language Models (V-LLMs). Our system introduces two novel contributions over traditional IEC approaches: (1) the use of LLMs as intelligent mutation and crossover operators that perform directed transformations of code and prompts, informed by implicit domain expertise; and (2) the use of visual LLMs as automated evaluators and feedback providers, not only to select offspring for breeding but also to generate detailed, expert-level assessments that steer the evolutionary trajectory. This system is instantiated in two domains of high creative complexity: landing-page design and generative visual art. In both domains, LLM-driven transformations generate high-quality, functional, and aesthetically pleasing outputs, while visual LLMs act as domain-aware critics, iteratively refining designs toward optimal forms. Through extensive experiments and qualitative assessments, we demonstrate that our system significantly outperforms traditional IEC and stochastic design pipelines in terms of innovation, convergence speed, and quality of output. This work offers a blueprint for future autonomous design systems, revealing the untapped potential of V-LLMs in evolution-inspired creative processes.

ABSTRACT. In a non-generative approach to artificial intelligence in an artistic practice, this work looks at mapping processing data from different artificial neural networks (NNs) onto sound and visuals. One aim of this practice-based piece of research is to offer insights into how these ubiquitous, yet notoriously unintelligible algorithms operate, sometimes by exposing the audience to that very unintelligibility. The other is to use these vast amounts of abstract data as a blank canvas for audiovisual artworks. At the heart of the whole work is a link and cross fertilization between the use of sounds and visuals aesthetically associated with errors and digital malfunction, and the use of actual ‘waste’ data (from NN training), which acts as a trace of their operation. We look at different kinds of NNs, and at various in-training data-streams: from the easily apprehensible output of generative networks, through various performance metrics, to the highly-abstracted activation and weight values in and between hidden layers. We believe these pieces uncover some interesting features about the evolution of GANs’ outputs in training and the consistent patterns they follow in the latent space, illustrate the potential for errors (mode collapse, noise, etc.) in those networks, and outline the inherently abstract nature of weight changes during an MLP’s training.

ABSTRACT. The selection mechanism is one of the crucial parts of any genetic programming algorithm, which determines its efficiency. In this paper 23 different variants of selection schemes for genetic programming are compared, including tournament, lexicase, and Friedman ranking-based, where first and second parent can be chosen differently. Four environmental selection, also known as replacement schemes are compared, including elitism, pairwise comparison, fitness sorting-based and Friedman ranking-based. The genetic programming had two populations with different replacement schemes. The comparisons are made for standard and down-sampled lexicase selection on 96 regression datasets. The analysis of the results shows that some of the typically applied population management schemes are suboptimal and can be easily changed to more efficient ones.

ABSTRACT. While crossover is a critical and often indispensable component in other forms of Genetic Programming, such as Linear- and Tree-based, it has consistently been claimed that it deteriorates search performance in CGP. As a result, a mutation-alone (1 + λ) evolutionary strategy has become the canonical approach for CGP. Although several operators have been developed that demonstrate an increased performance over the canonical method, a general solution to the problem is still lacking. In this paper, we compare basic crossover methods, namely one-point and uniform, to variants in which nodes are “preserved,” including the subgraph crossover developed by Roman Kalkreuth, the difference being that when “node preservation” is active, crossover is not allowed to break apart instructions. We also compare a node mutation operator to the traditional point mutation; the former simply replaces an entire node with a new one. We find that node preservation in both mutation and crossover improves search using symbolic regression benchmark problems, moving the field towards a general solution to CGP crossover.

ABSTRACT. Genetic programming has become a powerful tool for solving symbolic regression and similar problems, yet its inner workings often remain a mystery, prompting the need for techniques that can reveal and explain its evolutionary behaviour. Therefore, in this study, we adapt search trajectory networks (STNs) to analyse and visualise the search process of tree-based genetic programming (GP) for symbolic regression problems. STNs are graph-based models of the dynamics of search and evolutionary algorithms, where nodes represent search states and edges represent transitions between consecutive states. As network models, they can be analysed and visualised with the rich graph theory and complex networks toolsets. We consider three STN models: a standard search space or genotype model, where nodes are syntax trees, and two semantic or phenotype models, where nodes encode regression output vectors. The two semantic models differ in the partitioning technique used to group sets of output vectors into STN nodes: (i) standard hypercubes, and (ii) a newly proposed approach using an advanced clustering method. Standard tree-based GP is contrasted against a recent variant of geometric semantic GP. Our analysis reveals that STNs quantitatively and visually capture the essential search and performance differences between these two GP paradigms. The newly proposed semantic STNs produce compact and interpretable models, which can scale up to real-world symbolic regression problems.

ABSTRACT. Model selection criteria in symbolic regression have used training error and complexity measures to balance underfitting against overfitting. Standard metrics like AIC, BIC, and MDL combine these elements in different ways, based on the intuition that high complexity may allow erratic data-fitting, including overfitting to noise. However, these criteria do not directly evaluate model behaviour in extrapolation regions or measure prediction stability under input perturbations. We introduce hybrid model selection criteria that extend information-theoretic measures with explicit penalties for aspects of bad model behaviour not captured by standard metrics. Our approach combines base metrics (MSE, AIC, BIC, MDL, PSM) with three additional components: (1) extrapolation divergence, which measures prediction changes between interpolation and extrapolation regions; (2) interpolation sensitivity, which quantifies prediction variance under perturbations in dense data regions; (3) extrapolation sensitivity, which evaluates prediction stability in sparse regions. Experiments across 17 regression datasets from the PMLB repository test various weighting of these measures using both single-variable and multiple-variable perturbation methods. By testing different weight configurations, we aim to identify formulas that select models with better overall performance, balancing training fit, complexity, extrapolation stability, and prediction sensitivity.

ABSTRACT. In recent years, the world has changed considerably thanks to spectacular advances in the field of generative artificial intelligence. However, it is undeniable that this new technology requires ever-increasing amounts of scarce resources. Against a backdrop of shortages in computing power, there has been an increase in work on expensive optimization (EO), particularly in the field of evolutionary computation (EC) when solving real-world problems. In this article, we attempt to define EO, or at least characterize it. We also demonstrate its growing importance. We then summarize the many strategies integrated into EC that have been proposed in the literature for dealing with EO. However, we also demonstrate that these strategies primarily focuses on reducing execution time. Furthermore, we emphasise that this cost should not be the only factor to be taken into account. Based on these considerations, we depict several research directions that should be investigated for expensive evolutionary computation.

ABSTRACT. Evolutionary computation (EC) suffers from a lack of standardized open-source frameworks, the need of domain-specific configurations for every new problem, and missing general guidelines for choosing appropriate solution methods. Unlike other areas of artificial intelligence (AI), such as machine learning, which provide ready-to-use programming libraries with reusable components, EC often requires algorithms to be implemented from scratch for each new problem. In this paper, a vision for the next stage of EC is proposed, introducing an architecture for an agentic generative AI system designed to unify and automate EC processes, enabling higher levels of integration, autonomy, and self-organization. This represents a step toward a new paradigm for interacting with agentic and generative AI, supporting dynamic, multi-agent collaboration in which problem modeling, algorithmic design, and efficient computation are seamlessly integrated across the entire EC ecosystem. Ultimately, this could reduce the entry barrier and foster broader adoption of EC.

ABSTRACT. Benchmarking has driven scientific progress in Evolutionary Computation, yet current practices fall short of real-world needs. Widely used synthetic suites such as BBOB and CEC isolate algorithmic phenomena but poorly reflect the structure, constraints, and information limitations of continuous and mixed-integer optimization problems in practice. This disconnect leads to the misuse of benchmarking suites for competitions, automated algorithm selection, and industrial decision-making, despite these suites being designed for different purposes.

We identify key gaps in current benchmarking practices and tooling, including limited availability of real-world-inspired problems, missing high-level features, and challenges in multi-objective and noisy settings. We propose a vision centered on curated real-world-inspired benchmarks, practitioner-accessible feature spaces and community-maintained performance databases. Real progress requires coordinated effort: A living benchmarking ecosystem that evolves with real-world insights and supports both scientific understanding and industrial use.

ABSTRACT. Automated algorithm design is entering a new phase: Large Language Models can now generate full optimisation (meta)heuristics, explore vast design spaces and adapt through iterative feedback. Yet this rapid progress is largely performance-driven and opaque. Current LLM-based approaches rarely reveal why a generated algorithm works, which components matter or how design choices relate to underlying problem structures. This paper argues that the next breakthrough will come not from more automation, but from coupling automation with understanding from systematic benchmarking. We outline a vision for explainable automated algorithm design, built on three pillars: (i) LLM-driven discovery of algorithmic variants, (ii) explainable benchmarking that attributes performance to components and hyperparameters and (iii) problem-class descriptors that connect algorithm behaviour to landscape structure. Together, these elements form a closed knowledge loop in which discovery, explanation and generalisation reinforce each other. We argue that this integration will shift the field from blind search to interpretable, class-specific algorithm design, accelerating progress while producing reusable scientific insight into when and why optimisation strategies succeed.

ABSTRACT. This paper presents observations and an assessment of the Evolutionary Computation (EC) ecosystem from a sociotechnical perspective. While EC exhibits a strong progressive trajectory in its technical dimensions, this paper examines the field's social dimensions through four people-centred axes: the application-focused practitioner perspective, EC education, the need to revisit natural evolution for deeper foundational abstractions, and the status of negative-results publishing in EC. These observations and assessments are supported by a profiling of major EC conferences and events. The analysis identifies both strengths and opportunities in practitioner support, pedagogical coherence, sustained efforts to return to biological sources, and dedicated venues for negative results. Actionable suggestions aimed at strengthening EC's social dimension are also presented to support a more socially aware evolution of the EC community.

ABSTRACT. Heating, Ventilation, and Air Conditioning (HVAC) systems represent one of the largest sources of energy consumption in residential buildings, a demand expected to increase as climate control becomes more widespread in developing regions. In recent years, data-driven approaches have been increasingly applied to enhance building energy efficiency. Yet, many of these models operate as opaque black boxes, limiting user trust and understanding. To address this challenge, interpretability has emerged as a critical aspect of intelligent control systems. In this work, we propose an Interpretable Artificial Intelligence (IAI) framework that integrates Reinforcement Learning (RL) and Evolutionary Computation within a Federated Learning (FL) setting. This approach enables the evolution of transparent, Decision Tree (DT)-based control policies using distributed data from multiple buildings in diverse climatic regions, while preserving data privacy. We evaluate the proposed method in a simulation-based optimization environment using Sinergym, demonstrating its effectiveness in achieving a balanced trade-off between energy consumption, occupant comfort, and interpretability.

ABSTRACT. A widespread transition to electric buses (eBuses) is essential for reducing CO2 emissions and achieving global environmental goals. However, this transition demands a significant upfront investment to strategically place the charging infrastructure required to operate an eBus transportation network. A key challenge in this transition is addressing range anxiety, a concern that arises due to the limited range of electric buses on a single charge. Range anxiety impacts the operational reliability, making it crucial to deploy charging infrastructure in carefully chosen locations rather than installing it without a clear strategy. To tackle these challenges, this paper proposes a local search framework aimed at minimising the deployment costs of charging infrastructure for a fleet of eBuses. The framework incorporates operational constraints and supports the shared use of charging infrastructure across the transportation network. To evaluate our framework, we employ a realistic cost model from the literature that takes into account multiple factors, including the hardware and installation costs, and apply it to two Irish cities

ABSTRACT. This article presents an evolutionary algorithm for the automated design of sewer networks, a relevant problem for public health in modern cities. The approach follow the condominial model, a low-cost design strategy applicable in low- and middle-income scenarios. The evolutionary algorithm is meant to be integrated into the Sanitation and Basic Infrastructure Design (SaniBID) system by the Inter-American Development Bank for the design and evaluation of urban sanitation projects. A realistic hydraulic model is considered in the problem formulation. The proposed method is evaluated on real-world scenarios from Montevideo, Uruguay, showing consistent improvements of up to 45\% in total installation cost compared to the minimum slope greedy heuristic currently applied for sewer network design in the SaniBID system.

ABSTRACT. The rapid growth of mass tourism poses significant challenges to visitor experience and places substantial environmental and operational pressures on heritage sites and urban areas. This study examines the optimization of individual tourist trips and evaluates their broader urban sustainability impact, using Perugia, a historical city in central Italy, as a case study. In particular, we propose a recommendation system for tourist itineraries that, once configured with data on points of interest, takes tourist preferences as input and generates an optimized trip. Each trip is produced by solving a specifically formulated combinatorial optimization problem aimed at maximizing tourist satisfaction. We demonstrate that the problem is NP-hard and we propose three heuristic algorithms to address it. While optimizing individual trips, algorithm executions also update relevant city-wide state variables, managing queues at the different points of interest, which are then used to assess cumulative urban impact. Experiments were conducted as simulations across multiple scenarios with varying numbers of tourists and different arrival patterns. The results compared favorably with a baseline method and demonstrated that well-engineered micro-level trip decisions can positively influence city-wide outcomes

ABSTRACT. Solving a combinatorial optimization problem involves finding a consistent assignment of values to variables that satisfy a set of constraints while optimizing some objectives. However, modeling the problem can be a tedious task that requires strong expertise. In this context, modeling the Nurse Scheduling Problem (NSP) in healthcare units may be achieved actively through an expert or passively using historical data. Furthermore, passive modeling may be done manually or automatically. Manual modeling involves examining historical data and eliciting the problem’s constraints. However, this process may be tedious given the massive amount of available data. Therefore, automatic modeling may be considered to overcome this challenge in practice (especially in case of incomplete knowledge about the constraints and preferences). Constraints can be learned implicitly using machine learning methods or explicitly using constraint learners. In this paper, we propose two modeling approaches to learning the NSP; the first approach aims at learning a constraint network, represented as a Constraint Satisfaction Problem (CSP), and the second consists of approximating the NSP constraints and preferences from historical data. More precisely, we tackle the following research questions: ``How can we explicitly learn a CSP model for the NSP given historical data?'' and ``How can we approximate the constraints and preferences of the NSP using historical data?''

ABSTRACT. The growing deployment of Electric Vehicles (EVs) presents critical operational challenges, particularly in assigning limited charging infrastructure while ensuring user satisfaction. This work addresses the Electric Vehicle Charging Scheduling Problem (EVCSP) to minimize total tardiness, i.e., the cumulative delay beyond EVs’ intended departure times. We consider a centralized, coordinated, and offline scheduling setting, involving heterogeneous, non-preemptive, fixed-capacity chargers. A Mixed-Integer Linear Programming (MILP) formulation is first developed to model charger assignments and scheduling constraints. Given the computational complexity for solving large instances to optimality, we introduce a greedy, priority-based heuristic and a General Variable Neighborhood Search (GVNS) metaheuristic. The heuristic uses a tailored urgency index to assign vehicles rapidly, while GVNS explores multiple neighborhoods and local optima to enhance solution quality. Both approaches are benchmarked using synthetic scenarios with different fleet sizes, time-window tightness, and charger configurations. Results show that GVNS consistently outperforms the heuristic and the exact model in larger instances, achieving near-optimal solutions within short runtimes. The proposed hybrid methodology provides a scalable and practical solution for EV fleet operators seeking efficient, real-time-ready scheduling under infrastructure constraints.

ABSTRACT. The Oven Scheduling Problem (OSP) is a complex combinatorial optimization task characterized by parallel batching and heterogeneous ovens operating under tight production constraints. Existing methods, such as constraint programming and local search, are effective but often limited by computational cost or restricted search diversity. This paper introduces CP-MEME, a hybrid (1+1)-Evolutionary Framework that combines the constraint-based feasibility reasoning of constraint programming with the adaptive exploration capabilities of evolutionary computation. In the initialization phase, a CP-SAT model rapidly produces a feasible and near-optimal baseline schedule. The subsequent phase employs a single-individual (1+1)-Evolution Strategy enhanced by adaptive local search to intensify promising solutions while preserving feasibility. When search stagnation is detected, a penalty-guided perturbation mechanism executes batch swaps to diversify the search trajectory and escape local optima. Experimental evaluations on benchmark OSP datasets demonstrate that CP-MEME outperforms standalone CP-SAT and state-of-the-art local search methods, achieving consistent improvements in both solution quality and computational robustness.

ABSTRACT. When optimizing the performance of science codes, they must be adapted to the target hardware every time a new system is bought. This task is typically done manually and is both time-consuming and error-prone. Recently, machine learning has been applied to code optimization; however, most existing approaches are black-box models that provide no guarantees regarding the correctness or legality of the transformations they produce. TADASHI is a Python library that bridges the gap between compiler technologies and machine learning. It enables machine learning experts with little to no compiler experience to transform and optimize their code while generating a list of transformations that clearly explain each modification step by step. In the original paper, the authors proposed applying machine learning to TADASHI for code optimization, making the process automatic, and provided simple illustrative examples. This work introduces EvoTADASHI, a genetic programming–based approach that evolves a list of transformations to be used by TADASHI. We compare our results with both the heuristic approach proposed in the reference paper and our own implementation of beam search. The results show that EvoTADASHI and beam search achieve comparable performance, both with G-Mean speedups of 1.86x, while the latter requires twice as many evaluations to reach a solution. Furthermore, EvoTADASHI’s results improve to a 2.04x speedup when the heuristic approach is used to bootstrap the evolutionary process.

ABSTRACT. Selecting the most suitable algorithm for a given problem instance remains a challenging task, particularly in online or dynamic environments where problem characteristics evolve over time. Relying solely on instantaneous performance metrics can result in a reactive and unstable behaviour, often leading to suboptimal algorithm switching. This paper introduces a computationally efficient approach for aggregating an algorithm’s performance across multiple problem instances, that is fairly immune to erratic variations in instance features. Inspired by features inherent to Reinforcement Learning (RL), this technique, encapsulates rewards and penalties into a latent yield, that in turn triggers exploitation and exploration, consequently resulting in adaptive algorithm switching. The proposed technique employs island models, inspired by Genetic Algorithms, to facilitate parallel exploration and performance exchanges among algorithm populations inhabiting local repertoires. Experimental evaluations on sorting algorithms and robotic obstacle-avoidance tasks demonstrate the feasibility and effectiveness of the approach, highlighting its potential in domains where adaptive algorithm selection is critical.

ABSTRACT. Transformer models impose severe computational and memory demands on modern hardware accelerators, making efficient workload scheduling a key bottleneck for energy- and latency-constrained deployment. Mapping multi-head self-attention onto systolic-array architectures with hierarchical on-chip memory requires joint optimization of compute ordering, array assignment, and memory traffic. Accurate evaluation of such mappings must therefore be memory-aware, as traditional compute-only analytical models overlook bandwidth limits, buffer contention, and data-reuse effects, leading to unrealistic performance estimates. In this work, we introduce a simulation-driven scheduling framework that couples black-box optimizers with a high-level, memory-aware simulator to explore the combinatorial space of mapping Transformer workloads on systolic-array accelerators. The scheduler iteratively refines candidate execution plans using performance feedback from TransInferSim, which provides cycle-level latency and energy estimates through detailed modeling of compute and memory components. Multiple black-box optimizers are compared within this loop, revealing how memory-aware evaluation influences convergence dynamics and the relative performance of candidate schedules compared to naive compute-only estimation. Our experiments showed a 8.49% improvement in the latency of MHSA computation when using our memory-aware scheduling framework compared to the compute-only optimized baseline. This demonstrates the critical importance of memory-aware evaluation for efficient real-world deployment of Transformer workloads on modern hardware accelerators.

ABSTRACT. Graph-based genetic programming (GGP) encompasses representations that evolve modular graphs or computer programs with multiple inputs and outputs, making it well suited for addressing complex real-world problems. To fully exploit this potential, variation operators need to capture and preserve structural dependencies within program graphs. The gene-pool optimal mixing evolutionary algorithm (GOMEA) is a model-based evolutionary algorithm whose strength lies in learning and exploiting such dependencies, making it a natural candidate for GGP. In this work, we investigate the integration of GOMEA with GGP. We first validate the approach on symbolic regression (SR) benchmarks, with both single and multiple outputs, where GOMEA consistently matches or outperforms a standard genetic algorithm (GA). Then, we apply GOMEA to continuous control tasks—an important application domain for GGP—and find it often struggles compared to the GA. We hypothesize that this limitation arises from the difficulty of identifying and exploiting meaningful dependencies in the inherently chaotic and high-dimensional landscapes of control problems. Thus, our findings call for further studies to improve dependency-learning mechanisms for complex, dynamic domains.

ABSTRACT. Over the past few decades, machine learning has been widely used to learn complex tasks. Reinforcement Learning (RL), inspired by human behavior, is a great example, as it involves developing specific behaviours for specific tasks. To further challenge algorithms, Multi-Task RL (MTRL) environments have been introduced, requiring a single model to learn multiple behaviors. In this work, we present a new benchmark based on the MuJoCo Half Cheetah from Gymnasium. This benchmark features five distinct obstacles that are randomly positioned in front of the agent, each of which demands a unique behavior. This benchmark serves as a use case for Multi-Action TPG (MATPG), an extension of Tangled Program Graph (TPG) designed for MTRL environments with continuous control. MATPG aggregates MAPLE agents, which are Genetic Programming (GP) agents tailored for continuous control, and creates a control flow to activate them. Although MATPG initially showed no improvement over MAPLE agents on classic MuJoCo tasks, our experiments demonstrate its superiority in this multi-task use case when combined with lexicase selection. Furthermore, we examine the interpretability of the evolved graph, revealing that the decision flow of the model is fully interpretable.

ABSTRACT. In this work, we propose an Active Learning (AL) framework for Cartesian Genetic Programming (CGP) to evolve data-efficient programs for biomedical image segmentation. Active Learning enables the dynamic selection of training data by identifying the most informative images to include during evolution. To further reduce computational cost, we introduce a filtering stage that selects a diverse subset of candidate images using Principal Component Analysis (PCA) before uncertainty evaluation. We also investigate several strategies for selecting the initial training data: the most typical image, cluster-based, and random, and analyse their influence on convergence and model diversity. Our results show that filtering substantially reduces computational overhead while preserving data diversity, and that the choice of both initialization and uncertainty metrics significantly impacts convergence speed and overall performance. The proposed approach demonstrates that integrating Active Learning and filtering into CGP leads to faster convergence, improved performance, and more efficient utilization of training data.

ABSTRACT. With the need for enhanced interpretability, Cartesian Genetic Programming (CGP) has proven to be a promising alternative to widely used black-box approaches in fields such as image processing or program synthesis. Yet, the approach often fails to reach state of the art performance due to fitness landscape characteristics which make it difficult to efficiently explore the search space. We extend Multimodal Adaptive Graph Evolution (MAGE), a multi-type variant of CGP, with Monte Carlo Tree Search (MCTS), a search heuristic known for its capability to overcome these difficulties. MAGE-MCTS outperforms the classic 1+λ Evolutionary Algorithm, and shows a significant increase in performance, particularly for problems where 1+λ guided search tends to get trapped in local optima or neutral fitness plateaus. We show how MCTS-guided search can explore deceptive search paths and reach optimal solutions by traversing low-fitness search regions. Our contribution illustrates the need for problem- and fitness landscape-specific search strategies in CGP with the aim to efficiently explore the search space.

ABSTRACT. The use of semantic awareness in Genetic Programming (GP) has attracted growing attention in recent years. In this framework, the semantics of a program are defined as the vector of outputs it produces on the training cases and loss, commonly used as a fitness measure in GP for supervised learning, corresponds to a distance between the program’s semantic vector and the target vector from the dataset. This formulation has an implicit consequence: programs performing well in densely pop- ulated output regions tend to receive higher fitness scores and are thus favored during evolution. Conversely, poor performance in sparse regions may go unnoticed, as these areas are underrepresented and contribute little to the overall fitness evaluation. This work aims to mitigate this imbalance by promoting the survival of programs whose performance is more uniform across the entire space of output values. To this end, we introduce the Extended Semantics Operator (ExtenSO), a mechanism that probabilistically expands an individual’s semantic vector by gener- ating artificial output values through oversampling. These synthetic val- ues are added in underrepresented areas of the output space, producing a more homogeneous semantic distribution, spanning more diverse out- put values. Experimental results show that GP enhanced with ExtenSO achieves performance comparable to, or better than, standard GP, while generating smaller individuals. ExtenSO acts as a semantic regularization operator, reducing overfitting and encouraging smoother, more balanced behavior across the output domain. These findings confirm the relevance of semantic awareness in GP and introduce a simple, effective way to exploit it for greater accuracy and robustness.

ABSTRACT. Semantic complementarity among trees plays an important role in Multi-Tree Genetic Programming (MTGP) based feature construction. It enables trees to capture diverse yet relevant information, thereby reduce multicollinearity among constructed features and improve generalization. However, this complementarity is often constrained by syntax-driven initialization, which overlooks semantic diversity and leads to trees with highly correlated outputs in the initial population, and by conventional genetic operators that disregard semantic relationships and disrupt useful complementarities among trees. To overcome these limitations, this paper proposes a semantic clustering initialization method that promotes semantic complementarity by assembling individuals with trees originate from different semantic clusters. In addition, a component replacement crossover is introduced to preserve this complementarity by exchanging semantically related trees during evolution. Experiments on 12 regression datasets show that the proposed MTGP framework consistently improves predictive accuracy and convergence over MTGP baselines, confirming the benefits of incorporating semantic clustering and semantically informed variation operators.

ABSTRACT. Evolutionary Computation (EC) and Swarm Intelligence (SI) include some of the most successful meta-heuristics for global optimization. In practical applications, selecting the most suitable EC or SI meta-heuristics as well as the proper setting of its hyper-parameters is usually expensive in terms of computational resources and might be troublesome for novice users. To overcome these downsides, we propose a novel global optimization method, called Hybrid CMA-PSO (HyCAPS), which does not demand any complicated procedure for the identification of the hyper-parameters values as it exploits two settings-free algorithms: covariance matrix adaptation evolution strategy (CMA-ES) from EC, and Fuzzy Self-Tuning Particle Swarm Optimization (FST-PSO) from SI. HyCAPS is structured as a two-islands algorithm that evolves one population running CMA-ES and another running FST-PSO, which cooperate by exchanging the fittest individuals at a given frequency. We show that HyCAPS outperforms four optimization methods -- FST-PSO, jSO, NL-SHADE-RSP, NL-SHADE-LBC -- and competes with CMA-ES and SHADE-ILS on many high-dimensional, continuous, and bounded optimization problems taken from the IEEE CEC'05 and CEC'17 benchmark suites. Our results also highlight that the cooperation between the islands limits the risk of getting stuck in local optima and enhances the search space exploration. We believe that HyCAPS will represent an advantageous meta-heuristics in real-case scenarios where no information about the optimization problem is generally available.

ABSTRACT. Optimization problems in practical domains, such as engineering design, are increasingly characterized by a larger number of variables, complex behaviors, and multiple constraints, thereby making the tasks significantly more challenging. Particle Swarm Optimization (PSO) has been widely used for various types of problems due to its simplicity and effectiveness even in non-convex settings; however, its performance degrades as the number of decision variables increases. In this study, we present a novel PSO variant, named DPM-PSO, that integrates three key enhancements to extend its effectiveness to high-dimensional problems: (1) dynamic learning of variable interactions, (2) particle updates based on the learned interactions, and (3) a mutation mechanism to maintain diversity in complex search spaces. These enhancements synergistically reinforce the optimization efficiency of PSO. We evaluated the proposed algorithm on four benchmark functions and compared its performance with several state-of-the-art PSO variants. The results demonstrate that the DPM-PSO consistently achieves highly competitive search performance in high-dimensional settings.

ABSTRACT. This paper explores the evolutionary dynamics of information flow and control strategies in a simulated soft robotic system. By integrating a Genetic Algorithm with Reinforcement Learning, leveraging the Ray RLlib library together with the EvoGym simulation framework, we investigate how structural cost penalization, Lamarckian policy inheritance, and the initial population's conditions influence the emergence of centralized versus decentralized control architectures, as well as the input activation patterns, in Voxel-Based Soft Robots (VSRs). Our findings demonstrate that introducing a cost based on the distance over which control and input signals propagate significantly alters the search space. In cost-free settings, hybrid centralized/decentralized control architectures and unstructured input preferences tend to emerge. Conversely, the presence of signal propagation costs promotes decentralized control strategies and more selective input usage. Moreover, the cost constraint acts as a regularizer, improving average task performance across the population while limiting the occurrence of exceptionally high-performing individuals; the best task performers overall were achieved in the absence of such costs. We also observed that policy inheritance significantly affects fitness and task performance, while varying the initial population's conditions has minimal impact on the evolutionary outcomes.

ABSTRACT. Electron Ionization (EI) libraries enable fast identification of unknowns from their mass spectra, but classical dot–product–style search falters when the observed spectrum is a superposition and chro- matographic support is unavailable. We present a deconvolution workflow that relies solely on the EI–MS spectrum, requiring no external informa- tion. The method was developed with an industrial laboratory partner and evaluated on synthetic and real spectra. The method employs a two–stage strategy. First, a diagnostic–aware whitelist screens the library by rewarding agreement on strong, informative ions while penalizing un- supported intensity and maximizing over small integer m/z shifts to ab- sorb minor misalignment. Second, a PSO-driven greedy builder assembles a sparse mixture from the shortlist, allowing a bounded per–component power stretch to accommodate modest intensity variability, followed by a brief joint refinement. We also introduce a similarity ordering of the library (cosine on log normalized spectra with optimal–leaf ordering) so that local neighbor scans probe look–alike references without alter- ing reported indices. The fitness combines global NMSE, peak–weighted NMSE, and a spectral–angle term to emphasize diagnostic ions while preserving overall shape. On instrument–like synthetic mixtures and pre- liminary checks on laboratory and field EI spectra, the approach yields high–recall reconstructions; while over–selection persists in our tests, this is a reasonable trade-off given the size and redundancy of EI libraries, and preferable to missing true constituents, with settings that are straight- forward to apply in practice.

ABSTRACT. A central premise in evolutionary biology is that individual variation creates information asymmetries, which in turn facilitate the emergence of hierarchical organization. Using an agent-based model (ABM), we investigate the minimal conditions under which hierarchies arise in dynamic multi-agent systems—specifically, the roles of {\em initial heterogeneity} among agents and {\em mutation amplitude} across generations. The degree of hirarchical organisation is quantified through trophic analysis. We hypothesize that even small individual differences can be amplified into stable hierarchies through repeated local interactions involving reproduction, competition, and cooperation. Our results support this hypothesis. Hierarchical structures are more sensitive to mutation amplitude than to initial heterogeneity. Repeated experiments show that stable hierarchies reliably emerge only when initial heterogeneity remains low while mutation amplitude is sufficiently high. This framework captures both the dynamics and resilience of emergent social order, offering a quantitative account of how structured inequality can evolve from initially homogeneous populations. The findings have broader implications for evolutionary theory, collective organization, and artificial life research.

ABSTRACT. Cervical cancer screening demands methods that are both scalable and temporally aware. We present a hybrid framework that couples an agent-based layer with a Hidden Markov Model (HMM) for state transitions across two follow-up windows (S1: 0-6 months; S2: 6-12 months) and a neural network that produces calibrated state probabilities and 6-month forecasts for auxiliary variables. The agents constrain plausible moves among diagnostic categories; the HMM formalizes persistence, progression, and regression, and links latent states to observations; finally, the neural network synthesizes these signals into risk-ready pre dictions aligned with routine follow-up. On test data, state-classification accuracy reached 0.979 in S1 and 1.000 in S2, with ROC–AUC ≈ 1 in both windows. Furthermore, not only does the predicted correlation structure closely match that of the observed data, but the odds-ratio analyses are also clinically coherent. These findings show that integrating the Agent+HMM structure with neural predictors yields stable, temporally consistent, and clinically coherent performance, offering a practical path toward risk-based decision support in cervical screening workflows.

ABSTRACT. The rising global incidence of skin cancer, particularly Basal Cell Carcinoma (BCC), necessitates the development of automated diagnostic systems that are both accurate and interpretable for clinical adoption. While Deep Learning approaches have shown promise in medical image analysis, their ``black box'' nature remains a significant barrier to clinical trust and deployment. To address these challenges, this study introduces CoDED (Co-evolutionary Descriptor Engine for Detection), a genetic programming method that enhances existing automated BCC detection through two synergistic advancements. First, the method replaces traditional binary encoding with evolved Local Ternary Patterns (LTP), enabling richer textural feature extraction through adaptive dual-threshold mechanisms. Second, a novel Discriminative Localisation Mapping (DLM) technique provides pixel-level interpretability, revealing the model's diagnostic rationale through saliency visualisations. Comprehensive experimental evaluation on clinical dermoscopy images demonstrates that CoDED achieves a balanced accuracy of 79.57\%, significantly outperforming benchmark methods. The DLM analysis unveils distinct activation topologies; sparse, localised patterns for BCC lesions versus dense, widespread activations for healthy tissue, providing clinically meaningful insights into the model's decision-making process. These findings establish CoDED as a powerful and transparent diagnostic tool that bridges the gap between AI capability and clinical interpretability, offering a trustworthy foundation for computer-aided diagnosis in dermatology.

ABSTRACT. Traditional cryptography, rooted in problems, e.g., integer factorisation or discrete log, is inevitably vulnerable to a fully operational quantum computer. Although it remains an engineering frontier, the looming threat extends to encrypted data stored today, which could be decrypted in the future with quantum capabilities. To safeguard against this eventuality, the backbone of the modern quantum-safe cryptography is the Shortest Vector Problem (SVP). We enhance Laarhoven’s treatment of Ajtai et al.’s sieving as a genetic algorithm (GA) for the SVP by incorporating domain-informed SVP representation and crossover while naturally extending application to the module lattices.

ABSTRACT. Physics-informed machine learning algorithms increasingly rely on methods that balance interpretability, predictive accuracy, and computational efficiency. This paper highlights the central role of differential equation discovery, particularly of ensembles of equations, in understanding and improving learned dynamics. Using ensembles, we discover the uncertainty in the governing equations to describe complex dynamical systems. To further refine this understanding, we consider simple ensembles that could be considered as a special case of stochastic differential equations (SDEs) and more complex ensembles built with the Bayesian Network. We use Sobol indices to assess the influence of conditional distributions and thus the impact of a more sophisticated ensembling approach.

ABSTRACT. Boolean functions with strong cryptographic properties, such as high nonlinearity and algebraic degree, are important for the security of stream and block ciphers. These functions can be designed using algebraic constructions or metaheuristics. This paper examines the use of Evolutionary Algorithms (EAs) to evolve homogeneous bent Boolean functions, that is, functions whose algebraic normal form contains only monomials of the same degree and that are maximally nonlinear. We introduce the notion of density of homogeneous bent functions, facilitating the algorithmic design that results in finding quadratic and cubic bent functions in different numbers of variables.

ABSTRACT. This work explores the use of evolving assistive controllers to improve user performance in target-pointing tasks within human- computer interaction. We investigate two assistance paradigms: (1) vec- torial controllers that directly predict target-oriented displacement vec- tors (2) agent-based controllers trained via Reinforcement Learning. Each is implemented using either artificial gene regulatory networks or arti- ficial neural networks. artificial gene regulatory networks were evolved using a genetic algorithm, while artificial neural network agents were optimized via backpropagation in the first case and a Soft Actor–Critic algorithm in the second case. Thirteen participants performed a point- ing task using one of three input modalities (primary hand, secondary hand, and joystick). We have evaluated the impact of each controller on throughput, workload, and perceived control. Results show that as- sistive controllers improved participants’ performances, particularly the agent-based controller consistently improved throughput and reduced workload, indicating faster and more efficient cursor control. These find- ings demonstrate the feasibility of learning-based controllers to enhance cursor precision and reduce effort in human-computer interaction tasks. Future work will address target-agnostic operation by integrating gaze tracking or computer-vision approaches to enable more autonomous and adaptive assistive control systems.

ABSTRACT. This paper examines the feasibility of using a quantum-inspired genetic algorithm (QGA) to solve the multi-objective job-shop scheduling problem (JSSP) on classical computers. Since most real-world problems addressed by evolutionary algorithms have extremely large search spaces, exploring and integrating quantum computing concepts into evolutionary algorithms to solve these problems presents an intriguing avenue for research. While quantum computing has potential for solving combinatorial optimization problems, most existing quantum solutions require quantum hardware, which is still limited experimentally. To address this, we investigate how quantum-inspired mechanisms, simulated on classical computers, can enhance traditional evolutionary algorithms for combinatorial optimization problems. Experimental results show that incorporating quantum-inspired components to a classical genetic algorithm can enhance solution quality and convergence in tackling the multi-objective JSSP.

ABSTRACT. This paper evaluates the impact of hybridizing a multi-modal Genetic Algorithm with a Graph Neural Network for timetabling optimization. The Graph Neural Network was used to encapsulate general knowledge for improving the quality of the schedules, while the Genetic Algorithm integrated the deep learning model as an operator to guide the solution search towards optimality. Initially, both components of the hybrid technique were designed, developed, and optimized to solve the tackled task independently. Multiple experiments were conducted over Staff Rostering, a well-known timetabling problem, to compare the proposed hybridization with the standalone optimized versions of the Genetic Algorithm and Graph Neural Network. The experimental results show that the proposed hybridization brings statistically significant improvements in the time efficiency and solution quality metrics in contrast to the standalone methods. To the best of our knowledge, this paper proposes the first hybridization of a Genetic Algorithm with a Graph Neural Network for solving timetabling problems.

ABSTRACT. Water distribution networks (WDNs), as critical infrastructure, face growing cyber-attack risks. While Industry 4.0 initiatives motivate the deployment of edge technologies for monitoring, efficient detection on distributed edge devices requires methods that reduce computational overhead without sacrificing classifier performance. We propose a hybrid evolutionary-machine learning (ML) approach that constructs probabilistic surrogate models of trained ML classifiers using an Estimation-of-Distribution Algorithm (EDA). Our het-EDA algorithm supports the analysis of heterogeneous network communication features, extending r-UMDA for categorical and PBIL-C for continuous data. It evolves class-wise surrogates which optimize the original classifier’s output scores. Inference is performed via lightweight log-likelihood evaluation, making the method suitable for resource-constrained edge devices. Experiments on the WDN a Hardware-in-the-Loop dataset show that the surrogates preserve portions of the original decision boundaries, while highlighting limitations due to feature independence assumptions. These results demonstrate the potential of EDA-based probabilistic surrogates for efficient edge ML inference and motivate the development of more expressive EDAs suited for complex industrial network data.

ABSTRACT. Attention head pruning in transformers has been explored to reduce computational complexity while maintaining strong relative performance. Therefore, this work compares traditional pruning criteria with metaheuristic-based approaches. Instead of relying on expensive training from scratch or iterative fine-tuning, it aims to investigate an alternative strategy for navigating the intricate dependencies between attention heads within each self-attention module of a transformer. Specifically, this study applies structured pruning to attention head pruning in Vision Transformers and examines how various algorithms can assess the relative importance of different attention heads. The most effective approach identified in this study was based on Simulated Annealing combining accuracy and computational effort into a single normalized objective that guided the optimization toward balanced pruning solutions. This method preserved 95.64\% of the original model performance while reducing the number of attention heads by 55\% compared to the unpruned model.