ABSTRACT. The Complete Vocal Technique (CVT), a school of singing developed in the past decades by Cathrin Sadolin et al. CTV groups vocals into so called vocal modes, namely Neutral, Curbing, Overdrive and Edge. Knowledge of the desired vocal mode can be helpful for singing students. Automatic classification of vocal modes can be important for technology-assisted singing teaching. Previously, automatic classification of vocal modes has been attempted without major success, potentially due to a lack of data. Therefore, we recorded a novel dataset made up of sustained vowels recorded from four singers, three of which professional singers with more than five years of CVT-experience. The dataset covers the entire vocal range of the subjects -- 3752 unique samples in total. By using four microphones for recording, this way offering a natural data augmentation, the dataset consists of more than 13000 samples combined. An annotation was created using three CVT-experienced annotators, each providing an individual annotation. The merged annotation as well as the three individual annotations come with the published dataset. Additionally, we provide some baseline classification results.

Best balanced accuracy across a 5-fold cross validation of 81.3\,\% was achieved with a ResNet18. The dataset can be downloaded under (review link only) \url{https://zenodo.org/records/17600363?token=eyJhbGciOiJIUzUxMiIsImlhdCI6MTc2MzAzOTU2MSwiZXhwIjoxNzY5ODE3NTk5fQ.eyJpZCI6IjI5MzQyZDEwLWY0MTQtNGU0NC05MzY1LTBmMmZlNjBlYmRmMSIsImRhdGEiOnt9LCJyYW5kb20iOiIzNWM3MDA3M2ExZGE2ZmM3NmI3NDc3MWRkMzBlZTdhNCJ9.9iK3UXRr-Gwc1b0YhM1kZIy308Th3aMD-ylasyou1MVwBeMR_S9JhnmIAR_p-36Rcp08n4pGyd_Zp1Vm-E0JIA}.

ABSTRACT. This article proposes the application of generative models for the autonomous musical exploration of astronomical data to be used both in creative applications and scientific outreach. It is focused on the catalog of variable sources observed by the Optical Monitoring Camera (OMC) onboard the International Gamma-Ray Astrophysics Laboratory (INTEGRAL) hosted by the Spanish Virtual Observatory (SVO). The work describes a methodology, based on Box Least Squares (BLS) periodograms and LSTM networks, for the exploration of astronomical catalogs through sound and music. The proposal explores the representation of light flux time-series through musical notes calculated from the main periodic signal detected in the represented light curves. The resulting notes are cross-matched with a generative composition created from a selection of musical pieces byWolfgang Amadeus Mozart. The autonomous composition Mozart’s sky is presented as a proof of concept that provides real sky examples of the creative and informative possibilities behind this approach.

ABSTRACT. Creative coding and real-time shader programming are at the forefront of interactive digital art, enabling artists, designers, and enthusiasts to produce mesmerizing, complex visual effects that respond to real-time stimuli such as sound or user interaction. However, despite the rich potential of tools like GLSL, the steep learning curve and requirement for programming fluency pose substantial barriers for newcomers and even experienced artists who may not have a technical background. In this paper, we present AI Co-Artist, a novel interactive system that harnesses the capabilities of large language models (LLMs), specifically GPT-4, to support the iterative evolution and refinement of GLSL shaders through a user-friendly, visually-driven interface. Drawing inspiration from the user-guided evolutionary principles pioneered by the Picbreeder platform, our system empowers users to evolve shader art using intuitive interactions, without needing to write or understand code. AI Co-Artist serves as both a creative companion and a technical assistant, allowing users to explore a vast generative design space of real-time visual art. Through comprehensive evaluations, including structured user studies and qualitative feedback, we demonstrate that AI Co-Artist significantly reduces the technical threshold for shader creation, enhances creative outcomes, and supports a wide range of users in producing professional-quality visual effects. Furthermore, we argue that this paradigm is broadly generalizable. By leveraging the dual strengths of LLMs—semantic understanding and program synthesis—our method can be applied to diverse creative domains, including website layout generation, architectural visualizations, product prototyping, and infographics. We also explore whether human curators in the interactive process could be replaced or augmented with multimodal vision-language models acting as autonomous aesthetic judges to allow closed-loop evolution.

ABSTRACT. Estimation-of-distribution algorithms (EDAs) have shown strong performance in multi-valued optimization. Self-adjusting parameter control has been used in evolutionary algorithms to improve convergence speed and stability; however, to the best of our knowledge, these mechanisms have not yet been applied within the EDA framework. In this work, we introduce the self-adjusting multi-valued compact genetic algorithm (SM-cGA) and the compact genetic neuroevolution algorithm (cGNA). Both algorithms integrate success-based parameter control into the multi-valued compact genetic algorithm (r-cGA) and neural network optimization, respectively, allowing them to automatically adjust their parameters during the search. We evaluate the SM-cGA on G-OneMax function and the cGNA on four geometric benchmark problems. The results show that both algorithms consistently outperform static parameter settings and classical neuroevolution methods, such as the (1+1) neuroevolution algorithm with local mutation. We also provide theoretical runtime analyzes of the cGNA on two geometric benchmark problems, supporting the experimental results.

ABSTRACT. Gray-box optimization leverages the information available about the mathematical structure of an optimization problem in order to design efficient search operators. The DRILS algorithm (Deterministic Recombination with Iterated Local Search) has been used to solve Adjacent NK Landscape problems with up to one million variables to optimality. The deterministic recombination operator is Partition Crossover. In this paper we use DRILS for bi-objective optimization where the target functions are Random and Adjacent NK Landscapes. Our approach introduces a new way of vectoring the target objective functions. This bi-objective implementation of DRILS is both extremely fast and highly effective when compared to general implementations of NSGA-II and MOEA/D, and it naturally extends to multi-objective implementations.

ABSTRACT. The Two Level p-Median Location Problem is an NP-hard combinatorial problem. Algorithm Selection Methods have been extensively studied in the literature, to solve such problems. These studies have emphasized the importance of analyzing fitness landscape characteristics. In this paper we focus on ruggedness and neutrality features. First, we propose estimating both features using various machine learning models on reproducible instances. The tests show that models based on Random Forest and Gradient Boosting produce the best results overall with a coefficient of determination R2 exceeding 0.9 for both features. Then, we introduce a reverse model, FLEM-Rev based on an evolutionary algorithm to generate instances that match one or both of the targeted landscape features. FLEM-Rev either uses the RF predictors or GB predictors to predict features in the evaluation function. Two types of exercise are tested. The first type involves features whose values are close to those in the training set, and the second type involve features whose values fall within unexplored zones of the training-test space. The results show that our reverse model performed well in both cases when the features are considered separately.

ABSTRACT. Classical competitive coevolution has long utilized subjective measurements of historical progress and post hoc competition to quantify results due to intractable search spaces. This has contributed to divergent perspectives and disjointed literature, which pose an obstacle to new researchers and practitioners. In response to a point of confusion in literature, this work investigates the potential for the Hall of Fame technique to produce deceptive fitness approximation effects in competitive coevolution and draws a through line between disconnected literature. Using a parameterized number game with variable dimensions of intransitivity and a predator-prey meta-game, we conduct an investigation of fitness approximation during competitive coevolution using fully-enumerated search spaces to establish objective ground truth. We observe that fitness approximations during competitive coevolution can result in deceptive solution ranking that produces breakdowns in elitism and global search progress, as theorized in literature, and objectively quantify this effect for the first time. We identify that these effects are directly related to the presence of intransitivity and show that deceptive effects increase with Hall of Fame use, but also with increasing dimensions of intransitivity and high selective pressure in survival selection.

ABSTRACT. This paper compares training strategies for evolving autonomous agents (bots) designed to play the deck-building card videogame, Tales of Tribute, using the open-source \textit{Scripts of Tribute} framework. The developed agent employs a greedy strategy guided by a heuristic evaluation function, optimizing its 20 defining numerical weights using an Evolutionary Strategy. Three different training modes have been used: Fixed (against static reference bots), Coevolution (internal peer-to-peer competition), and a Hybrid approach (combining segments of fixed and coevolutionary training). The influence of the Hall of Fame memory technique across these modes is also analysed.

ABSTRACT. Dota 2 represents a challenging multiplayer online battle arena video game in which players attempt to define strategies for `hero' characters, where each hero has different abilities. Spell casting is one of the more complex skill based behaviours for a hero to acquire. In this work, we assume a transfer learning approach for developing spell casting behaviours for Invoker, one of the game's heroes. We show that a suite of Invoker agents can be identified by developing a set of 9 quantifiable behaviour classes and 4 spell usage behaviours. All the top performing Invoker agents demonstrate a subset of the behaviour classes with above normal frequency without ignoring any of the others. Conversely, mid performing Invokers tended to rely on specific behaviour classes while ignoring others. The worst performing agents either hand poor strategies or were very specialized at beating particular opponent agents.