Automated planning is a prominent area of Artificial Intelligence, and an important component for intelligent autonomous agents. A critical aspect of domain-independent planning is the domain model, that encodes a formal representation of domain knowledge needed to reason about a given problem. Despite the crucial role of domain models in automated planning, there is a lack of tools supporting the knowledge engineering process by comparing different versions of the models, in particular, determining and highlighting differences the models have.In this paper, we formalise the notions of structural equivalence and strong equivalence of domain models, and, on top of these, we introduce a novel concept of similarity of domain models. To measure the similarity of two models, we introduce a directed graph representation of lifted domain models that allows the formulation of the domain model similarity problem as a variant of the graph edit distance problem. We propose an Answer Set Programming approach to optimally solve the domain model similarity problem, that identifies the minimum number of modifications the models need to become strongly equivalent, and we demonstrate the capabilities of the approach on a range of benchmark models.
Successful human-agent teaming relies on an agent being able to understand instructions given by a (human) principal. In many cases, an instruction may be incomplete or ambiguous. In such cases, the agent must infer the unspoken intentions from their shared context, that is, it must exercise the principal’s Theory of Mind (ToM) and infer the mental states of its principal. We consider the prospects of effective human-agent collaboration using large language models (LLMs). To assess ToM in a dynamic, goal-oriented, and collaborative environment, we introduce a novel task, Instruction Inference, in which an agent assists a principal in reaching a goal by interpreting incomplete or ambiguous instructions.We present Tomcat, an LLM-based agent, designed to exhibit ToM reasoning in interpreting and responding to the principal’s instructions. We implemented two variants of Tomcat. One, dubbed Fs-CoT (Fs for few-shot, CoT for chain-of-thought), is based on a small number of examples demonstrating the requisite structured reasoning. One, dubbed CP (commonsense prompt), relies on commonsense knowledge and information about the problem. We realized both variants of Tomcat on three leading LLMs, namely, GPT-4o, DeepSeek-R1, and Gemma-3-27B. To evaluate the effectiveness of Tomcat, we conducted a study with 52 human participants in which we provided participants with the same information as the CP variant. We computed intent accuracy, action optimality, and planning optimality to measure the ToM capabilities of Tomcat and our study participants. We found that Tomcat with Fs-CoT, particularly with GPT-4o and DeepSeek-R1, achieves performance comparable to the human participants, underscoring its ToM potential for human-agent collaboration.
Effective coordination is the cornerstone of success in cooperative Multi-Agent Reinforcement Learning (MARL), where agents work together to achieve a collective goal. However, most existing methods for modeling agent coordination rely primarily on proximity between agents’ features or on incorporating heterogeneity at the behavioral/policy level, while overlooking the valuable diversity present at the observation level. In this paper, we introduce Diametric Coordination Graphs (DiaCoG), a novel framework that dynamically models implicit coordination of agents by integrating both consistency (shared similarities) and discrepancy (unique differences) from agents’ observations, offering a richer understanding of inter-agent relationships. Accordingly, we propose two approaches, DiaCoG-DE and DiaCoG-CE, implementing this framework under two MARL architectures of actor-critic networks, Centralized Training and Decentralized Execution (CTDE) and Centralized Training and Centralized Execution (CTCE), respectively. Through theoretical analyses based on information-theoretic measures, we show that DiaCoG offers superior expressiveness over consistency-based methods for value estimation and action selection. Empirically, we evaluate the proposed methods in the Predator-Prey and Traffic Junction environments, where they outperform baselines in terms of both final average returns/success rates and convergence speed across diverse scenarios. To further showcase the adaptability of DiaCoG across different MARL paradigms, we integrate it with a state-of-the-art value-based approach and compare it against several representative and state-of-the-art methods to demonstrate the enhanced coordination performance by DiaCoG in the StarCraft II Multi-Agent Challenge environment.
Answer Set Programming (Asp) is used in many real-world applications as an easy-to-use and efficient modeling paradigm. In Asp, rules (containing variables) are instantiated against facts to derive conclusions (i.e., grounding the program). However, a large number of variables and facts (more precisely: domain values) leads to the well-known “grounding bottleneck”: the variables in a rule are instantiated by every combination of domain values, leading to potentially infeasibly large ground programs. This is especially problematic for rules with large rule bodies containing many variables, as standard grounding approaches cause an exponential blowup in the total number of variables in a rule. This paper presents a novel method to overcome this obstacle. In contrast to the standard approach, our technique “decouples” the body atoms, i.e., the instantiation handles each body atom separately. In particular, our procedure translates a non-ground tight program to a ground disjunctive program that is exponential only in the maximum predicate arity, and is thus polynomial if this arity is fixed to a constant. We extend this to non-ground, normal programs, and ultimately use similar ideas to translate non-ground, disjunctive programs to ground epistemic logic programs—all while decoupling the body atoms. We demonstrate the feasibility of this new method experimentally by comparing it to standard Asp technology in terms of grounding size, grounding time and total runtime.
Rationalization is a framework that aims to build self-explanatory NLP models by extracting a subset of human-intelligible pieces of their input texts. It involves a cooperative game where a generator (i.e., selector) selects the most human-intelligible parts of the input as the rationale, followed by a predictor that makes predictions based on these selected rationales. In this study, we aim to equip the rationalization framework with the potential to mine causal rationales from a purely probabilistic perspective, without involving strong assumptions implicit in graphical causal models (e.g., conditional independence and the ignorability of unobservable confounders). The two obstacles that prevent the conventional rationalization framework from selecting causal rationales are spurious correlation and degeneration, where the predictor overfits the spurious or meaningless rationale candidates solely selected by the not-yet well-trained generator and in turn deteriorates the generator. In this paper, we propose a simple yet effective method named MGR (multi-generator based rationalization) to simultaneously solve the two problems, and thus extract the causal rationales. The key idea of MGR is to employ multiple generators such that the occurrence stability of causal parts is improved and more meaningful pieces are delivered to the predictor. Through experiments on six text classification datasets and one graph classification dataset using three network architectures (GRUs, BERT, and GIN), we show that our method significantly outperforms existing rationalization methods. We also compare our method with a representative LLM (LLaMA-3.1-8B-Instruct) and find that our simple method achieves comparable results to those of the LLM.
We propose Fast Concept-based Counterfactual Explanations (FCCE) for generating counterfactual explanations for CNN image classification models in real time. While counterfactual explanations are intuitive for humans, existing vision methods often face a trade-off between semantic interpretability and computational efficiency, limiting their widespread adoption in interactive decision support systems. FCCE addresses this challenge by operating in a semantic concept space and deriving a closed-form analytical counterfactual solution, with the remaining computational overhead dominated by standard feature extraction. To evaluate practical utility, we conducted a between-subject human study with 197 participants on a fine-grained model-failure identification task on the CUB dataset. Participants achieved 89.88% decision accuracy with FCCE, outperforming CVE (50.89%), ACE (68.47%), and ICE (82.95%), and reported higher perceived understanding, confidence, and satisfaction. Complementing the behavioral results, we assess fidelity and computational efficiency on CUB, MNIST, and ImageNet. Crucially, we demonstrate that FCCE counterfactuals remain highly faithful when mapped back to the original non-linear CNN, achieving a Relative Logit Gap below 0.3%, thereby providing strong empirical evidence for the fidelity of our counterfactual explanations. Finally, FCCE computes the analytical counterfactual solution in under 10−5 seconds on CPU, with an end-to-end local explanation overhead of approximately 10−2 seconds per instance. This is orders of magnitude faster than both optimization-based baselines and the representative source-conditioned diffusion-based generative counterfactual pipeline used as a latency reference. Overall, FCCE reconciles semantic interpretability with real-time performance and improves human decision-making about model reliability.
Active distribution networks are undergoing a rapid transition driven by the increasing penetration of distributed resources, including renewable generation, demand response programs, and electric vehicles, which introduces new technical and operational challenges for day-ahead network management. Coordinating conventional control devices together with emerging distributed flexibility leads to a nonlinear, nonconvex, mixed-discrete, and multi-period optimization problem with high computational complexity. To address this challenge, this paper proposes an intelligent solution methodology based on genetic algorithms, reinforcement learning, and dynamic programming for short-term operational planning of distribution systems. The methodology adopts a two-level structure. At the aggregator level, electric-vehicle charging is optimally scheduled to flatten the loading profile of secondary distribution transformers. At the distribution system operator level, controllable resources, including demand response actions, OLTC tap positions, capacitor banks, and network reconfiguration switches, are coordinated to minimize operating costs associated with energy losses, voltage violations, branch congestion, and controllable-resource use. To enhance search efficiency and reduce dependence on manual parameter tuning, a tabular Q-learning agent is embedded into the genetic algorithm to adaptively control selection, crossover, and mutation strategies during execution. The best hourly operating candidates obtained by the genetic algorithm are then linked through dynamic programming to determine the optimal day-ahead schedule while preserving temporal consistency. The proposed framework combines engineering-grade power-flow simulation with adaptive artificial-intelligence techniques, providing an effective and practical decision-support tool for modern distribution control centers.
The recent development of generative artificial intelligence has driven new forms of interaction between students, teachers, and technological systems within higher education. Among these tools, educational chatbots are positioned as conversational assistants capable of providing personalised support, facilitating access to academic information, and mediating learning processes in different educational environments. This paper analyses two experiences of implementation of educational chatbots at the National Technological University (UTN): TutorVirtualUTN, integrated in the virtual campus of the La Plata Regional Faculty (FRLP) using the ChatBotPress platform with GPT-3.5 Turbo, supporting 127 first-year engineering students in Algebra and Mathematical Analysis; and ProfeVirtual, deployed via the POE platform (Claude Instant model) as a digital twin of the teacher in Probability and Statistics, implemented across two cohorts (50 and 45 students) at the Avellaneda Regional Faculty (FRA). The study adopted a quantitative descriptive approach based on surveys applied during the 2024 and 2025 academic years. Results show comprehension rates improving from 56
This paper presents the development of a mobile application to support the identification of missing persons through facial recognition, integrating image processing services via Amazon Web Services (AWS). The system centralizes official data, including images and personal details, within a Spring Boot backend, PostgreSQL database, and Flutter mobile frontend. The facial recognition module compares registered and captured images, generating potential matches with similarity percentages. Functional tests validated the system’s integration and performance. The proposed solution provides technological support to institutions, enabling fast and scalable facial analysis through the integration of the Amazon Recognition cloud service, which employs pre-trained convolutional neural networks (CNNs).
This article documents a pedagogical co-creation experience between an expert professor and the Gemini Large Language Model (LLM) for the development of an assessment ecosystem for the Project Risk Management course. The core of the proposal is the application of the Condition-Risk-Consequence (CRC) methodology. We describe a progressive creation cycle, which transforms initial theoretical concepts into 15 high-fidelity assessment scenarios, situated within the Argentine macroeconomic context and quantified using precise financial variables. Based on this experience, the Layered Pedagogical Scaffolding Model (MAPC) is formally proposed, linking its structure to the Prompt Engineering best practices documented by Google. The model establishes an architecture where the quality of the AI response is directly proportional to the structuring of the context, role, and constraint layers. This methodological approach not only enhances the extreme personalization of learning but also ensures rigorous academic standards and professional relevance. The study concludes that the strategic use of LLMs like Gemini, guided by a clear scaffolding methodology such as MAPC, allows for overcoming the limitations of traditional assessment exercises. The ability to generate hyper-realistic and contextualized scenarios, with quantified financial variables, validates the potential of AI to raise the quality of higher education in technical and management fields.
Generative artificial intelligence (GenAI) is increasingly being explored to enhance learning experiences in virtual laboratories in engineering education. However, a systematic understanding of how GenAI integrates into virtual laboratories and how these approaches relate to pedagogical roles, learning outcomes, and implementation challenges remains limited. This scoping review analyses 14 studies of GenAI in virtual laboratories for higher engineering education. Based on the analysis, different GenAI integration approaches were identified: agent-based integration, API or service-based integration, tool or plugin-based integration, and external interface integration. Furthermore, these integrations were examined in relation to their pedagogical roles, learning outcomes, and implementation challenges. The findings indicate that GenAI can act as an intelligent tutor or learning tool to support multiple instructional functions in virtual laboratory environments, such as providing guidance, feedback, and interactive learning support. These roles contribute to improving students’ conceptual understanding, problem-solving skills, and engagement in laboratory learning. By mapping GenAI integration approaches with reported pedagogical roles and challenges, the study finds that the design of GenAI-integrated virtual laboratories should carefully balance automation, instructional support, and learner autonomy to achieve effective educational outcomes. These findings offer useful insights for both researchers and educators seeking to design effective GenAI-supported virtual laboratory learning environments.
Artificial intelligence and digital marketing are increasingly converging in healthcare organizations seeking to improve patient acquisition, retention, and follow-up. However, despite the growing interest in AI-enabled personalization, predictive analytics, chatbots, and automated engagement, empirical evidence in real-world private healthcare settings remains limited. This study examines the relationship between artificial intelligence and digital marketing in private healthcare organizations in Tacna, Peru. A quantitative, basic, correlational, non-experimental, and cross-sectional design was applied. Data were collected through a structured questionnaire administered to 110 private healthcare organizations registered in RENIPRESS, using a census approach for the eligible organizational frame. Artificial intelligence was measured through four dimensions: machine learning, cognitive intelligence, big data, and deep learning. Digital marketing was assessed through patient acquisition, patient retention, and patient follow-up. The inferential analysis used Spearman’s rho, considering the ordinal nature of the measurement scale and the distributional behavior of the variables. The results showed a positive and statistically significant relationship between artificial intelligence and digital marketing (rho = 0.562, p < 0.05). The findings suggest that AI adoption may strengthen data-driven marketing capabilities in private healthcare by supporting segmentation, personalization, automation, and patient relationship management. This study contributes empirical evidence from a regional Latin American healthcare context and offers practical implications for digitally transforming patient-centered marketing strategies.
Corruption in Peruvian public construction is often preceded by deficiencies in the Technical Project Dossier (TPD), the documentary package approved before bidding and execution. Existing tools quantify execution risk through FMEA or detect corruption from post-award procurement metadata, but they do not treat TPD approval as an ex-ante decision gate. This paper proposes AI+FMEA-ESG, a rule-based decision-support model that extends FMEA with a governance-weighted ESG modifier (see Eq. 1 below), bounded ±20 [ AdjRPN = RPN ×( 1 + ( ESG - 3)/10); Equation 1. Composite AdjRPN formula.; ] The model is anchored to SEACE/CONOSCE procurement data, a 233-case Basel Institute/CGR typology, and OSCE regulations. Results show that Phase F2 (Design and Development) is the principal intervention point: five defect modes reach Critical AdjRPN (≥270), with absent technical studies scoring highest (AdjRPN = 451). Seven Governance Solution Proposals translate these findings into auditable decision gates compatible with Peru’s digital procurement ecosystem. The revised version also discloses the scoring protocol, provides a worked example and sensitivity analysis, and sets out the current validation status explicitly.
Road and public-works heritage represents a strategic asset for sustainable development and public-value generation, yet its management across Latin America has historically been reactive rather than risk-informed. This paper presents the AI+FMEA model, which shifts Failure Mode and Effects Analysis (FMEA) from a static, design-stage exercise to a continuously updated, sensor-driven process. Building on a forensic review of 100 documented infrastructure failures (2020–2026), the study shows that 54
Communication barriers significantly marginalize millions of individuals with hearing and speech disabilities, limiting their access to essential services and social participation. This paper presents NeuroSign, an autonomous Edge-AI system designed to translate sign language into real-time speech and text through an innovative 3D holographic interface. Unlike cloud-dependent applications, NeuroSign utilizes a decentralized architecture powered by a Raspberry Pi 4 and an optimized TensorFlow Lite inference engine, ensuring data privacy and operational continuity in critical public infrastructure. The methodology integrates computer vision via MediaPipe Hands to map 21 spatial landmarks, which are then processed locally to trigger a “Pepper’s Ghost” holographic projection. Experimental results demonstrate a total system latency of 118 ms, well within the threshold for natural human interaction, and an average recognition accuracy of 70
This paper presents the design and empirical validation of Monty Lab, an interdisciplinary mechatronic ecosystem developed to address critical gaps in primary mathematics education within the Pan-American region. Integrating the VDI 2206 engineering standard with the Montessori pedagogical framework, the system comprises an AI-driven interactive robot, 3D-printed manipulative modules, and a structured activity book. The central processing unit, powered by a Raspberry Pi 5, utilizes computer vision via a HuskyLens AI camera to provide real-time, personalized feedback, creating a multimodal learning loop. A pilot study was conducted with a cohort of 30 third-grade students at I.E.P. Maryam School (Huancayo, Peru). Statistical analysis via a paired-samples t-test demonstrated a highly significant improvement in learning outcomes (t(29) = −9.448, p < .001), with mean proficiency scores rising from 12.60 to 15.53. Beyond academic performance, the project successfully integrated neurodivergent students, proving the efficacy of affective computing in inclusive education. By leveraging accessible, locally developed technology, Monty Lab serves as a scalable model for digital sovereignty, transforming engineering research into a strategic geopolitical asset for human capital development. The findings suggest that such AI-driven ecosystems are vital for modernizing teaching and fostering competitive talent in the regional AI-driven economy.
Large language models (LLMs) are rapidly transforming access to information and digital services across multiple sectors. However, these benefits remain unevenly distributed across linguistic and territorial contexts. Communities speaking indigenous and other low-resource languages often face a double barrier: limited digital connectivity and limited AI support in their own language. This paper presents a practical local LLM-RAG framework designed for underserved multilingual environments. The proposed approach combines local model inference, retrieval over a corpus of validated documentary sources, lightweight language mediation, and optional cloud augmentation when connectivity is available. Rather than focusing exclusively on benchmark scale, the framework prioritizes operational continuity, affordability, factual grounding, and equitable access. As an initial pilot case study, the paper uses Shawi, an indigenous language of the Peruvian Amazon, grounded in a curated corpus of 48 official educational and linguistic documents from Peru’s Ministry of Education, of which two priority resources (a Shawi–Spanish dictionary and an orthographic writing manual) are indexed for the pilot. A feasibility pilot involving 50 structured queries shows that the full framework raises the pilot useful-response rate from 28
Freight transportation along urban logistics corridors is exposed to many risk factors that lead to accident events which have a large economic, operational, social and safety impact. The urban mobility systems are growing in complexity and so there is the need for a predictive approach that can accommodate a wide range of unknown variables when predicting accidents in freight transport. To this end, we propose a multidimensional artificial intelligence framework for accident prediction in freight transportation with a human, operational, vehicular, environmental, technology, and infrastructure-based approach. We have developed a process of identifying and operationalizing variables, mathematical modeling, and machine learning to predict accidents from the comprehensive perspective to achieve this objective. To evaluate the technical feasibility of the proposed framework, we conducted a proof of concept implementation on a synthetic dataset of 10,000 simulated freight transportation operations. We created an accident occurrence with a logistic model which includes the risk factors and interaction effects. Two machine learning algorithms, Logistic Regression and Random Forest were trained and evaluated using the usual classification parameters: Accuracy, Precision, Recall, F1-score and AUC-ROC. The results showed good predictive power in the simulation environment with Logistic Regression with an AUC-ROC of 0.989 and Random Forest with an AUC-ROC of 0.940. We also explored the driving hours, digitalization level and vehicle condition in our multidimensional model. Though these findings are not empirical proof-of-concept in practice, they highlight the internal consistency and technical feasibility of how the proposed approach works. Such research will establish the basis for future work that integrates artificial intelligence, transportation safety and Logistics 4.0 principles with freight transportation data.
