ABSTRACT. The recent surge in the development of attention mechanisms has made it possible for language models (LMs) to produce text on par with humans. Unfortunately, the autoregressive nature of attention-based LM decoders inhibits them from attending to future information. In this paper, we first define the task of future conditioning: the process of motivating an LM decoder to condition on a future event. We then propose the Future Sight story generation pipeline for fine-tuning an attention-based LM decoder on the future conditioning task. We evaluate Future Sight against three datasets – Writing Prompts, ROCStories, and RecipeNLG. We measure the utility of Future Sight on text inter-diversity and entailment against a fine-tuned GPT-2 baseline. Both our novel automated entailment evaluator and human readers found texts generated by Future Sight to more sensibly lead up to a future event. Additionally, we observe a decrease in inter-diversity for batches of predictions with a fixed future event, demonstrating positive conditioning effects of the future. Our intention for the Future Sight pipeline is as a human-in-the-loop system where authors can propose their own future plot events to guide an LM as it drafts a story.

ABSTRACT. This study explores the application of Artificial Intelligence Generated Content (AIGC) in the creation of ink-wash animation, a unique artistic form deeply rooted in Chinese cultural heritage. Despite its profound cultural and aesthetic value, ink-wash animation has been constrained by labor-intensive production processes and lengthy development cycles, limiting its modern appeal. Through a detailed case study of Dragon Gate, a martial arts-themed short film, this research examines how AIGC can transform the creation of ink-wash animation by enhancing production efficiency and expanding creative possibilities. The study also analyzes the advantages and limitations of AIGC, focusing on its ability to preserve the traditional essence of Chinese ink painting while addressing technical and artistic challenges. By integrating human creativity with AI capabilities, this research highlights the potential of AIGC to reinvigorate ink-wash animation and lays the groundwork for its future development in a contemporary context.

ABSTRACT. We explored how people perceived animations created by artificial intelligence (AI)-driven motion capture, manual keyframe technique, and AI-driven motion capture with manual cleanup methods. We presented our participants with short, full-body animation clips created using the three methods. Participants rated the appeal and naturalness of the animations, and we asked them to discern the creation method. Results revealed differences in perceived appeal and naturalness between manually created animations and those generated through AI-based methods, with manual animations consistently rated higher in appeal and naturalness. However, participants could not discern creation methods regardless of animation experience level, demonstrating an accuracy equivalent to random guessing. The qualitative analysis highlighted diverse perspectives with negative and positive views on AI use, with the most mentioned theme being the importance of quality regardless of creation method. The overwhelming majority of participants asserted that the degree of automatization would influence participants’ perceived value and effort put into an animation. Still, this group did not show divergent ratings, nor did it affect their overall agreeableness towards using AI in creative fields. This study contributes insights into the intersection of animation and AI, informing creators about the effect of different creation methods on audience perceptions.

ABSTRACT. We propose new univariate Estimation-of-Distribution Algorithms that involve neuro-evolution processes for solving pseudo-Boolean optimization problems. We carried out experiments on instances of the Quadratic Binary Unconstrained Optimization. The algorithms automatically discovered demonstrate their competitiveness compared with classic EDAs.

ABSTRACT. Algorithm selection is crucial in the field of optimization, as no single algorithm performs perfectly across all types of optimization problems. Finding the best algorithm among a given set of algorithms for a given problem requires a detailed analysis of the problem's features. To do so, it is important to have a diverse set of benchmarking instances highlighting the difference in algorithms' performance. In this paper, we evolve diverse benchmarking instances for chance-constrained optimization problems that contain stochastic components characterized by their expected values and variances. These instances clearly differentiate the performance of two given algorithms. We introduce a $(\mu+1)~EA$ for feature-based diversity optimization to evolve such differentiating instances. We study the chance-constrained maximum coverage problem with stochastic weights on the vertices as an example of chance-constrained optimization problems. The experimental results demonstrate that our method successfully generates diverse instances based on different features while effectively distinguishing the performance between a pair of algorithms.

ABSTRACT. Variable Neighborhood Search (VNS) is a prominent metaheuristic for solving combinatorial optimization problems. While adaptive strategies have been explored to enhance VNS, identifying effective criteria for adaptive decision making remains a challenge. This paper introduces an adaptive VNS variant that leverages Fitness Landscape (FL) Analysis to guide neighborhood selection. The proposed approach employs machine learning to predict the most suitable neighborhood structure for each iteration based on computed FL features, with feature importance assessed using Shapley values from explainable artificial intelligence techniques.

The effectiveness of the proposed method is validated through experiments on benchmark instances of the Traveling Salesman Problem (TSP), demonstrating superior solution quality compared to traditional VNS and competitive performance relative to other metaheuristics. This study underscores the potential of landscape-aware optimization to enhance metaheuristic performance for complex combinatorial problems while providing valuable insights into the influence of key landscape features on adaptive search strategies.

ABSTRACT. As the demand for environmentally sustainable computing grows, understanding energy consumption in AI systems has become increasingly important. This paper explores how hysteretic phenomena and implementation choices affect the energy consumption of evolutionary algorithms (EAs). Specifically, we consider the case of running EAs in batch and show how back-to-back executions can put a significant strain on the underlying processing device, resulting in increased energy consumption. An experimental analysis indicates that the introduction of short pauses can alleviate this problem and reduce consumption by up to 9% in the considered benchmark. We also conduct a comparative analysis between two twin implementations of the same EA library in Java and C++, revealing that the latter scales better in terms of energy efficiency and running time, thus underpinning the importance of implementation decisions and best practices when aiming to optimize an algorithm's energy consumption.

ABSTRACT. Climate change and inefficient irrigation practices have underscored the necessity for sustainable water management in agriculture. This study focuses on fig trees (Ficus carica L.), a crop adapted to arid conditions, whose irrigation needs must be managed to optimize yield without overuse of water. Leveraging a combination of Artificial Intelligence (AI) techniques—Genetic Algorithms (GAs) and Neural Networks (NNs)—this research proposes a cost-effective method to predict water requirements using sensor data and climatic variables.

Experiments conducted at the Scientific and Technological Research Center of Extremadura (CICYTEX) assessed various sensor configurations and combinations of sensors with climatic data. The study compared models using all available inputs versus reduced sets selected via GAs. Results indicate that using only three sensors achieves a near-optimal predictive accuracy (MSE = 0.0205) compared to models utilizing all sensors (MSE = 0.0135). Similarly, the integration of climatic variables showed the feasibility of achieving high accuracy with fewer data points.

The results highlight the potential of hybrid AI models to minimize reliance on sensors while maintaining prediction accuracy, thereby optimizing irrigation to meet the specific needs of the crop. This approach represents a step toward sustainable agricultural practices by balancing resource efficiency and crop productivity.

ABSTRACT. Making software greener is a process that includes identifying the functions that consume the most energy, developing a methodology that can measure precisely that energy consumption and eventually measuring that energy under different design decisions and circumstances to be able to, eventually, produce best practices for minimizing said consumption. In this paper we are focusing on well-known floating-point fitness functions: some functions included in the black box optimization benchmark that cover all different types of functions under study. In general, these fitness functions will be the single operation that consumes the most energy; this is why in this paper we use them to test a methodology that is able to measure the energy consumed by their implementation in a low-level language C++. We test different single-element representations (single and double precision) as well as individual level representation (fixed size vs. variable size), drawing conclusions on the adequacy and accuracy of the methodology as well as which combination of the above elements would consume the least.