This paper summarizes the activities regarding the development of Artificial Intelligence (AI) for Sustainability conducted by the members of the AIIS (Artificial Intelligence and Intelligent Systems) node of the University “G. d'Annunzio" of Chieti-Pescara (Ud'A).
There is a growing recognition of AI’s own environmental footprint and this calls for AI researches being environmentally responsible and aligned with sustainability values. In fact, AI can serve as a powerful instrument for addressing environmental and climate issues, optimising decision-making, improving energy eficiency, and facilitating the transition to a more sustainable future.
lutions in green-Oriented GAMES (ECOGAMES) funded by PNRR Mission 4, line 1.3, funded by the European Union – NEXTGENERATIONEU, “Future Artificial Intelligence – FAIR” project - PE0000013, Spoke 9, CUP D23C24000210006. The project aims at bridging the fields of game theory and environmental sustainability with a specific focus on integrating energy-eficiency in systems formed by selfish agents.
and Abstraction for Explanation and Analysis (SMARTK) funded by PNRR Mission 4, line 1.3, funded by the European Union – NEXTGENERATIONEU, “Future Artificial Intelligence – FAIR” project - PE0000013, Spoke 9, CUP D23C24000220006. The project aims at advancing the fields of argumentation, abstraction, and automated reasoning in knowledge representation.
funded by INdAM-GNCS. The ASIA project proposes a novel combination of methodologies, techniques, and tools for argumentation analysis, machine learning, and social network analysis.
tems funded by INdAM-GNCS, CUP E53C22001930001. The project aims to explore existing literature and improve upon the methods for validating abstract argumentation frameworks that have been previously suggested, while also potentially introducing new approaches.
tation Theory funded by INdAM-GNCS, CUP E55F22000270001. The aim of the project is to extend formal debate verification approaches in argumentation theory using new abstraction and probability-based of the groups. Considering environmental sustainabilinterpretations. ity, teamwork among diferent stakeholders (individuals, Multi-objective Optimization of Digitally Manu- communities, businesses, and governments) is essential factured Earth Building Components supported by for achieving common goals such as reducing pollution Neural Networks (MUD-MADE) funded by PNRR and carbon emissions, addressing environmental and cliMission 4 - Component C2, Investment 1.1, PRIN, CUP mate issues or managing natural resources sustainably. D53D23020070001. The project aims to propose a novel A notable class of coalition formation games is that of artificial intelligence-supported workflow useful for de- hedonic games, introduced in [3], in which agents have signing raw earth building components produced with preferences over the set of all possible agent coalitions, digital manufacturing technology (i.e. 3D printing, and the utility of an agent merely depends on the compoRobotic arm or Laser cutter). sition of the coalition she belongs to. Work on hedonic games mainly studies the existence, computation and Fracture Risk evaluation in bone metastatic pa- performance of stable solutions, i.e., solutions where no tients by Artificial InteLligence (FRAIL) funded by agent or group of agents has interest in deviating from PNRR Mission 4 - Component C2, Investment 1.1, PRIN, the outcome, with respect to several notions of stabilCUP D53D23013760006. The project investigates whether ity such as Nash or strong Nash stability, core stability, AI can produce a reliable and explainable decision sup- individual stability and so on (see [4] for a nice survey). port system that can assist physicians in complex care Some members of the AIIS node of Ud’A are actively decisions regarding patients afected by bone metastases, working on coalition formation games [5, 6, 7, 8]. by making their treatment as accurate as possible.
3.3. A Concurrent Language for
Interacting Argumentative Agents We now describe the scientific work related to AI for Many AI applications aim to mimic human behavior Sustainability that is carried out from the researchers and reasoning to allow machines to emulate human-like of the Ud’A node of AIIS, with particular focus on the thoughts and actions. A significant challenge lies in proprojects listed in Section 2. viding mechanisms to formally articulate specific types of knowledge, allowing machines to use it to reason and in3.1. On Green Sustainability of Resource fer new insights. Modelling the behaviour of concurrent agents that interact and reason in a dynamic environment
Selection Games is a dificult task. It requires tools that can efectively In our interconnected world, increasingly dependent on capture diferent types of interactions, such as persuasion digital platforms and, at the same time, marked by grow- and deliberation, while helping agents make decisions ing concerns about environmental sustainability, a press- or reach agreements. Argumentation theory provides ing issue demanding timely and eficient solution is the formal frameworks for representing and evaluating inreduction of power consumption of Information Tech- teracting arguments. It is therefore important to define a nology (IT) devices (personal computers, data centers, language for modeling the interaction of concurrent and networks). It is highly anticipated, in fact, that their argumentative agents in a distributed system. This lancontribution to the annual electricity consumption in guage should allow the representation of diferent types 2030 will exceed 10% of the total demand [1]. Motivated of dialogues, describing the reasoning process employed by these considerations, we have studied in [2] a multi- by agents, thus making it a powerful tool for agent interagent system in which agents compete for the usage of action. power-consuming resources and are charged a cost pro- A member of the AIIS node of Ud’A is actively working on portional to their fair share of the power consumption, the definition of a Concurrent Language for Interacting by investigating the (in)eficiency of stable solutions. Argumentative Agents [9, 10, 11, 12]. 3.2. Coalition Formation Games 3.4. Abstraction and Explanation Teamwork and coalition or group formation is an impor- In computer science, abstraction is the standard tool used tant and widely investigated issue in computer science to model and study complex systems. For example, abresearch. In many economic, social and political situa- straction is used in model checking, analysis and veritions, individuals carry out activities in groups rather ifcation of software, neural network analysis or robust than by themselves. In these scenarios, it is of crucial learning. importance to consider the satisfaction of the members The AIIS node of Ud’A has a strong background in abstraction, in particular on abstract interpretation and its many applications [13, 14]. In this context, by exploiting the algorithms [15, 16] and the tools [17, 18] developed by the members of the AIIS node of Ud’A we are designing abstractions and tools for the analysis of the concurrent language for interacting argumentative agents discussed in the previous section [9, 10] and for the analysis of power consumption of software and neural networks. 3.5. Automated Reasoning for Verification provided a time series forecast of NO2 and CO given four meteorological parameters: (i) air pressure, (ii) relative humidity, (iii) average daily temperature, and (iv) wind speed [29]. We have also recently proposed a novel recurrent neural network-based system for tracking and forecasting the dispersal of air pollutants PM2.5 on building sites based on established environmental parameters [30]. Preliminary findings for predicting pollutants in spatial domains can also be found in [31].
This research is conducted in collaboration with the Mathematical Institute of the Serbian Academy of Sciences and Arts, the University of East Sarajevo, Italferr S.p.a. and the Atmospheric Physics and Chemistry Laboratory (UdAtmo - www.atmo.unich.it).
By exploiting this research, policy-makers and urban planners can develop more efective strategies for pollution control and urban planning: AI helps in reducing the health risks associated with air pollution and also contributes to the broader goal of sustainable urban development.
ferent programming languages, business processes, networks, and in general software systems, can be modeled as satisfiability problems for Constrained Horn Clauses (CHC) [19].
The AIIS node of Ud’A has worked on the development of methodologies for transforming CHCs and verifying their satisfiability, with applications in: ∙ verification of reachability properties for imperative programs manipulating arrays [20, 21] and Algebraic Data Types (e.g. lists and trees) [22, 23, 24]; ∙ verification of relational properties among pro- 3.7. Machine Learning Models for grams [25] (e.g. equivalence, functionality, injectivity, monotonicity, non-interference); Ecological Footprint Prediction ∙ verification of reachability and controllability proper- In order to comprehend the efects of human activity on ties of business processes defined using BPMN with time the ecosystem, this research direction studies predictive extensions [26]; models for ecological footprint, measuring the speed in ∙ generation of CHC verification conditions [ 27] based consuming resources and generating waste compared on the operational semantics of the programming lan- to the speed of nature in absorbing human’s waste and guage and the proof rules of the considered class of prop- generating resources. erties. In the recent time, we constructed and evaluated four
Focusing on sustainability, we envision potential ap- hybrid machine learning models (artificial neural netplications on the verification of power consumption for work, random forest regression and K-nearest neighbor programs, gas usage for smart contracts on a blockchain, regression) for predicting the total ecological footprint and on checking relational equivalence of systems, before of consumption from multiple energy inputs and populaand after power consumption optimization. tion number. The adopted energy inputs are: (1) natural gas, (2) coal, (3) oil, (4) wind, (5) solar photovoltaic, and 3.6. Machine Learning to Forecast (6) hydropower sources that are the main sources of enParticulate Matter and Trace Gas ergy consumption [32]. We also generated time series Emissions for Air Quality Assessment vector autoregression prediction models of the ecological footprint based on energy parameters [33].
This research, focusing on the use of particulate matter The AIIS node of Ud’A is conducting this research with (PM) data—particularly PM10 and PM2.5 concentrations, the Mathematical Institute of the Serbian Academy of but also nitrogen oxides (NO) and ozone (O3)—to eval- Sciences and Arts and with the University of Belgrade. uate air quality, employs advanced machine learning techniques including neural network models to analyze 3.8. Social Sustainability and LLMs and interpret continuous data from regional air quality monitoring stations [28]. According to the European Green Deal and to the Digital
Furthermore, the integration of meteorological param- Services Act, creating and supporting digital environeters into the machine learning models enhances the ments that are safe and supportive for all individuals is accuracy of predictions about air quality variations un- one the objectives that underlines the broader sustainder specific environmental conditions, demonstrating the able development goals. Combating online misogyny, in ability of AI to provide deeper insights into environmen- particular, aligns with Sustainable Development Goal 5 tal health impacts. To this respect, using a Nonlinear of the United Nations. While recent advances in GeneraAutoregressive Exogenous (NARX) neural network, we tive Pretrained Language Models have raised important questions about their safety and fairness [34], many re- In this research direction, the AIIS node of Ud’A is colsearchers are exploring their use to enforce online safety. laborating with the Mathematical Institute of the Serbian Generative LLMs are indeed flexible tools to implement Academy of Sciences and Arts. data analysis automation task, including hate speech detection [35, 36] and stance/polarity estimation [37]. 3.10. Forecasting Models for Climate
Our experiments in misogynistic content classification [38] show that a zero-shot GPT 3.5-based classifier out- Variables and Renewable Energy performs traditional Deep Learning methods like BERT, Accurately estimating energy production from renewwithout the need for large annotated training sets. Re- able energy sources is crucial for ensuring a reliable and sults indicate that majority voting among multiple AI consistent supply, aiding in the planning and managing annotators, each prompted diferently, is efective, yet of the power grid. Predicting their power generation highlights potential for further improvement. Specifi- is notably challenging due to their dynamic behaviour, cally, we are investigating CoT and Planning-like pat- influenced by factors like time, weather parameters and terns to improve recognition of more specific classes, as location. Beyond the inherent complexity of these phemisogynistic derogation, treatment, or counter-speech. nomena, the dificulties are further compounded by the
In a related line of research, the AIIS node of Ud’A is limited availability of local real-time data, particularly for collaborating in the RightNets PRIN project, leaded by short-term forecasts. Therefore, the use of forecasting University of Macerata and University of Sapienza, with models is essential to achieve this goal. In the case of the goal of enhancing Social Media monitoring dash- missing measured data, Regional Climate Models (RCMs) boards [39] with AI tools in order to more efectively represent a practical and valuable tool for describing the monitoring political and electoral debate on social media. climatology of places. The Fifth-Generation Mesoscale
Another research line focuses on utilizing textual data Model (MM5) is one of the most adopted among them. from social media, particularly during crises and natu- However, various works point out its tendency to underral disasters, to aid disaster responders. However, these /overestimate weather parameters. We proposed DL2F studies face challenges such as dealing with unstructured, [43], a powerful deep learning model that combines exnoisy, and ambiguous data, varying user credibility, ex- perimental data with MM5 information to address these pertise, and bias, and the overwhelming volume of data challenges and enhance the accuracy of forecasting progenerated during disasters. Leveraging state-of-the-art cedures. Our system forecasts four essential weather machine learning and AI techniques, we have proposed variables that influence solar power potential: Global a methodological framework for damage assessment on Horizontal Irradiance (GHI), temperature, atmospheric tweets related to Hurricane Ida that integrates textual pressure, and relative humidity. The model uses a time classification of social media data, spatial analysis, and series for each variable as input and the forecast provided visual analytics to provide rapid responses during natural by the MM5 system. Then, it generates new predictions disasters [40]. for each variable in the next 3 days. The model’s architecture is based on a set of GRU (Gated Recurrent Unit) 3.9. Machine Learning Models for neural networks. The new more precise predictions are Classification of Energy Consumption then adopted for calculating the electrical energy production of a photovoltaic cell [44].
global demographic and economic changes. Accordingly, 3.11. Energy-Eficient Deep Learning this research direction investigates on prediction models of energy consumption. Deep learning models, particularly those at the cutting
In the recent time, we considered the application of de- edge, can require significant computational resources mographic and economic features as predictor variables which translate into high energy consumption. This for energy consumption. In order to categorize energy not only increases the cost of training and deploying consumption levels from (i) gross domestic product, (ii) these models but also has a substantial environmental CO2 emissions, and (iii) total population, we investigated impact due to the carbon footprint associated with enthe usage of multiclass ensembles of support vector ma- ergy production. In this research direction, we developed chines and linear discriminant analysis [41]. GloNet [45], an eficient network architecture that can
Also, we applied multiple linear regression and multi- self-regulate its depth, and thus its computational needs, layer perceptron to forecast energy usage [42]. Energy according to the specific requirements of the task. This consumption was the dependent variable, whereas the enhances both sustainability and eficiency by reducing gross domestic product, population as a whole, and CO2 energy consumption and minimizing environmental imemissions were considered the predictor factors. pact. Furthermore, GloNet improves the accessibility of advanced models for users with limited computational resources, thereby contributing to the democratization of tion problem, J. Artif. Intell. Res. 72 (2021) 1215– deep learning. This is particularly impactful in resource- 1250. intensive fields such as reinforcement learning [ 46, 47], [7] M. Flammini, B. Kodric, G. Monaco, Q. Zhang, Stratmaking them more attainable and broadly usable. egyproof mechanisms for additively separable and fractional hedonic games, J. Artif. Intell. Res. 70 3.12. Health Sustainability (2021) 1253–1279. [8] V. Bilò, G. Monaco, L. Moscardelli, Hedonic games Some members of the Ud’A AIIS node also work in the with fixed-size coalitions, in: Proceedings of AAAI, ifeld of health sustainability where we have studied the 2022. role of AI and the impact of machine learning methodolo- [9] S. Bistarelli, M. C. Meo, C. Taticchi, Timed concurgies to analyze the Long COVID syndrome, from clinical rent language for argumentation with maximum presentation through diagnosis [48]. parallelism, J. Log. Comput. 33 (2023) 712–737.
A multi-omics approach integrating MRI-based ra- [10] S. Bistarelli, C. Taticchi, M. C. Meo, An interleaving diomics and metabolomics in rectal cancer treatment semantics of the timed concurrent language for arpredicts patient responses, potentially reducing unnec- gumentation to model debates and dialogue games, essary surgeries. A novel machine learning model using Theory Pract. Log. Program. 23 (2023) 1307–1333. pre-treatment MRI in locally advanced rectal cancer opti- [11] S. Bistarelli, M. C. Meo, C. Taticchi, On the role of mizes neoadjuvant treatments for better organ preserva- local arguments in the (timed) concurrent language tion. A radiomics-based model distinguishes COVID-19 for argumentation, in: Proceedings of AIˆ3, 2023. from other acute lung diseases using HRCT, aiding early [12] S. Bistarelli, M. C. Meo, C. Taticchi, Timed concurand accurate management decisions. These methods im- rent language for argumentation: An interleaving prove personalized medicine by enhancing predictive approach, in: Proceedings of PADL, 2022. accuracy and treatment sustainability [49, 50, 51]. [13] G. Amato, M. C. Meo, F. Scozzari, On collecting semantics for program analysis, Theoretical ComAcknowledgments puter Science 823 (2020) 1–25. [14] G. Amato, F. Scozzari, Observational completeness We acknowledge the support of: the PNRR MIUR project on abstract interpretation, Fundamenta InformatiFAIR - Future AI Research (PE00000013), Spoke 9 - Green- cae 106 (2011) 149–173. aware AI; GNCS and GNSAGA groups of INdAM; the [15] G. Amato, M. Parton, F. Scozzari, Discovering inPNRR PRIN project MUD-MADE - MUlti-objective opti- variants via Simple Component Analysis, Journal mization of Digitally MAnufactureD Earth building com- of Symbolic Computation 47 (2012) 1533–1560. ponents supported by neural networks (PNRR Mission 4 [16] G. Amato, F. Scozzari, H. Seidl, K. Apinis, V. Vojdani, - Component C2, Investment 1.1, PRIN). Eficiently intertwining widening and narrowing, Science of Computer Programming 120 (2016) 1–24. [17] G. Amato, F. Scozzari, The ScalaFix equation solver, References in: FM, 2023, pp. 142–159.
[18] G. Amato, F. Scozzari, JGMP: Java bindings and [1] J. Lorincz, A. Capone, J. Wu, Greener, energy- wrappers for the GMP library, SoftwareX 23 (2023). eficient and sustainable networks: State-of-the-art [19] E. De Angelis, F. Fioravanti, J. P. Gallagher, M. V. and new trends, Seonsors 19 (2019) 4864. Hermenegildo, A. Pettorossi, M. Proietti, Analysis [2] V. Bilò, M. Flammini, G. Monaco, L. Moscardelli, and transformation of constrained Horn clauses for C. Vinci, On green sustainability of resource selec- program verification, Theory and Practice of Logic tion games with equitable cost-sharing, in: Proc- Programming 22 (2022) 974–1042. cedings of AAMAS, 2024. [20] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proi[3] J. H. Dréze, J. Greenberg, Hedonic coalitions: Opti- etti, Program verification via iterated specialization, mality and stability, Econometrica 48 (1980) 987– Sci. of Comp. Programming 95 (2) (2014) 149–175. 1003. [21] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proi[4] H. Aziz, R. Savani, Hedonic games, in: Handbook etti, A rule-based verification strategy for array of Computational Social Choice, Cambridge Uni- manipulating programs, Fundamenta Informaticae versity Press, 2016, pp. 356–376. 140 (2015) 329–355. [5] G. Monaco, L. Moscardelli, Y. Velaj, Stable out- [22] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proicomes in modified fractional hedonic games, Auton. etti, Solving Horn clauses on inductive data types Agents Multi Agent Syst. 34 (2020) 4. without induction, Theory and Practice of Logic [6] M. Flammini, G. Monaco, L. Moscardelli, M. Shalom, Programming 18 (2018) 452–469.
S. Zaks, On the online coalition structure genera- [23] E. De Angelis, M. Proietti, F. Fioravanti, A. Pettorossi, Verifying catamorphism-based contracts [38] C. Morbidoni, A. Sarra, Can llms assist humans using constrained Horn clauses, Theory and Prac- in assessing online misogyny? experiments with tice of Logic Programming 22 (2022) 555–572. gpt-3.5, in: CEUR WS, volume 3571, 2023, p. 31–43. [24] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proi- [39] M. Mameli, M. Paolanti, C. Morbidoni, E. Frontoni, etti, Constrained horn clauses satisfiability via cata- A. Teti, Social media analytics system for action inmorphic abstractions, in: Proc. of LOPSTR, 2023. spection on social networks, Social Network Anal[25] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proi- ysis and Mining 12 (2022).
etti, Predicate pairing for program verification, [40] E. del Gobbo, B. Cafarelli, L. Ippoliti, L. Fontanella, Theory Pract. Log. Program. 18 (2018) 126–166. Mining social media data for damage assessment [26] E. De Angelis, F. Fioravanti, M. C. Meo, A. Pet- in environmental disasters, in: Proc. of IES, 2023. torossi, M. Proietti, Semantics and controllability [41] R. Janković, A. Amelio, Z. A. Ranjha, Classificaof time-aware business processes, Fundamenta In- tion of energy consumption in the balkans using formaticae 165 (2019) 205–244. ensemble learning methods, in: ICACS, 2019. [27] E. De Angelis, F. Fioravanti, A. Pettorossi, M. Proi- [42] R. Janković, A. Amelio, Comparing multietti, Semantics-based generation of verification layer perceptron and multiple regression models conditions via program specialization, Science of for predicting energy use in the balkans (2018).
Computer Programming 147 (2017) 78–108. arXiv:1810.11333. [28] C. Colangeli, S. Palermi, S. Bianco, E. Arufo, P. Chi- [43] L. Caroprese, M. Pierantozzi, C. Lops, S. Montelacchiaretta, P. Di Carlo, The relationship between pare, DL2F: A deep learning model for the local pm2.5 and pm10 in central italy: Application of ma- forecasting of renewable sources, Comput. Ind. Eng. chine learning model to segregate anthropogenic 187 (2024) 109785.
from natural sources, Atmosphere 13 (2022). [44] C. Lops, M. Pierantozzi, L. Caroprese, S. Montel[29] R. Janković, M. Cosović, A. Amelio, Time series pare, A deep learning approach for climate paramprediction of air pollutants : A case study for serbia, eter estimations and renewable energy sources, in: bosnia and herzegovina and italy, in: Proceedings Proceedings of BigData, 2023.
of INFOTEH, 2019. [45] A. D. Cecco, C. Metta, M. Fantozzi, F. Morandin, [30] M. Mastromatteo, A. Amelio, A deep learning ap- M. Parton, GloNets: Globally Connected Neural proach for predicting air pollutants on the construc- Networks, in: Proceedings of IDA, 2024. tion site, in: Proceedings of MIPRO, 2024. [46] F. Morandin, G. Amato, R. Gini, C. Metta, M. Parton, [31] A. Cucco, L. Ippoliti, N. Pronello, P. Valentini, C. Za- G. Pascutto, SAI a Sensible Artificial Intelligence ccardi, Gaussian processes and deep neural net- that plays Go, in: Proceedings of IJCNN, 2019. works for spatial prediction, in: Proc. of SIS, 2023. [47] F. Morandin, G. Amato, M. Fantozzi, R. Gini, [32] R. Janković, I. Mihajlović, N. Štrbac, A. Amelio, Ma- C. Metta, M. Parton, SAI: A sensible artificial intelchine learning models for ecological footprint pre- ligence that plays with handicap and targets high diction based on energy parameters, Neural Com- scores in 9× 9 go, in: Proceedings of ECAI, 2020. put. Appl. 33 (2021) 7073–7087. [48] I. Ahmad, A. Amelio, A. Merla, F. Scozzari, A survey [33] R. Janković, I. Mihajlović, A. Amelio, Time se- on the role of artificial intelligence in managing ries vector autoregression prediction of the ecolog- long covid, Frontiers in Artificial Intelligence 6 ical footprint based on energy parameters (2019). (2023).
arXiv:1910.11800. [49] A. Delli Pizzi, A. M. Chiarelli, P. Chiacchiaretta, [34] P. P. Liang, C. Wu, L.-P. Morency, R. Salakhutdinov, et al., Mri-based clinical-radiomics model preTowards understanding and mitigating social biases dicts tumor response before treatment in locally in language models, in: Proc. of Machine Learning advanced rectal cancer, Scientific Reports 11 (2021) Research, volume 139, 2021, p. 6565 – 6576. 5379. [35] F. M. Plaza-Del-Arco, D. Nozza, D. Hovy, Respectful [50] A. Delli Pizzi, A. M. Chiarelli, P. Chiacchiaretta, or toxic? using zero-shot learning with language et al., Radiomics-based machine learning diferentimodels to detect hate speech, in: Proc. of ACL, ates “ground-glass” opacities due to covid-19 from 2023. acute non-covid-19 lung disease, Scientific reports [36] N. Vishwamitra, K. Guo, F. T. Romit, I. Ondracek, 11 (2021) 17237.
L. Cheng, Z. Zhao, H. Hu, Moderating new waves [51] M. V. Mattoli, F. Cocciolillo, P. Chiacchiaretta, et al., of online hate with chain-of-thought reasoning in Combined 18f-fdg pet-ct markers in dementia with large language models, 2023. arXiv:2312.15099. lewy bodies, Alzheimer’s & Dementia: Diagnosis, [37] H. Xu, S. Vucetic, W. Yin, Openstance: Assessment & Disease Monitoring 15 (2023) e12515.
Real-world zero-shot stance detection, 2022. arXiv:2210.14299.