Ebook: PAIS 2022

This volume contains the proceedings of the Eleventh Conference on Prestigious Applications of Artificial Intelligence (PAIS 2022), held on July 25th, 2022 in Vienna, Austria, and co-located with the 31st International Joint Conference on Artificial Intelligence and the 25th European Conference on Artificial Intelligence (IJCAI-ECAI 2022).

PAIS 2022 received 18 full paper submissions and 4 extended abstract submissions, and accepted 10 full papers and 3 extended abstracts. The topics ranged from autonomous navigation, air traffic control and satellite management to the optimization of industrial processes and human-in-the-loop applications.

We would like to thank the authors for submitting their works to the conference, and the program committee members for their excellent and timely work in reviewing them. We would also like to thank the Artificial Intelligence Journal for having sponsored the event, and the IJCAI-ECAI 2022 organisers for having hosted PAIS as a satellite event of the conference and for having managed all the logistics related to its organisation.

Andrea Passerini and Thomas Schiex

Program Chairs, PAIS 2022

In the context of Earth observation constellations, we consider the problem of allocating orbit slots to clients requesting some ownership of orbit portions overflying desired regions on Earth. This problem arises prior to operational scheduling of observation tasks, in constellations where users can directly communicate with the satellites using their own ground stations. Observation scheduling in the exclusive slots is then delegated to the clients themselves. To perform the allocation of exclusive slots, we propose a two-level optimization approach, where the optimization process (led by either utilitarian or fair criterion) explores the solution space using a feasibility checker based on a constraint solver. We experimentally evaluate and analyze their performance on randomly generated order books and real constellation configurations.

Existing work on generating hints in Intelligent Tutoring Systems (ITS) focuses mostly on manual and non-personalized feedback. In this work, we explore automatically generated questions as personalized feedback in an ITS. Our personalized feedback can pinpoint correct and incorrect or missing phrases in student answers as well as guide them towards correct answer by asking a question in natural language. Our approach combines cause–effect analysis to break down student answers using text similarity-based NLP Transformer models to identify correct and incorrect or missing parts. We train a few-shot Neural Question Generation and Question Re-ranking models to show questions addressing components missing in the student’s answers which steers students towards the correct answer. Our model vastly outperforms both simple and strong baselines in terms of student learning gains by 45% and 35% respectively when tested in a real dialogue-based ITS. Finally, we show that our personalized corrective feedback system has the potential to improve Generative Question Answering systems.

A lot of technological advances depend on next-generation materials, such as graphene, which enables better electronics, to name but one example. Manufacturing such materials is often difficult, in particular, producing graphene at scale is an open problem. We apply state-of-the-art machine learning to optimize the production of laser-induced graphene, an established manufacturing method that has shown great promise. We demonstrate improvements over previous results in terms of the quality of the produced graphene from a variety of different precursor materials. We use Bayesian model-based optimization to quickly improve outcomes based on little initial data and show the robustness of our approach to different experimental conditions, tackling a small-data problem in contrast to the more common big-data applications of machine learning. We analyze the learned surrogate models with respect to the quality of their predictions and learned relationships that may be of interest to domain experts and improve our understanding of the processes governing laser-induced graphene production.

Humanoid robots have been successfully used in artistic research areas, and many works have studied and implemented systems for robotic dance. However, only few works take into account the human evaluation of these artistic outputs. This work makes a step in the direction of addressing the complex task of defining criteria for the evaluation of robotic dance performances. For this aim, in the context of a Master course on Fundamentals of Artificial Intelligence (AI), we have organized a challenge among our students and the winner is decided on the basis of a questionnaire we defined for robotic dance evaluation. In addition, we created a public dataset that maps the features of each choreography to the judgements provided by audience with different backgrounds on several evaluation targets. Then, we tested various Machine Learning models for predicting the audience evaluation, and we propose a choreography features importance analysis to help both human choreographers and AI algorithms to create dance performances with a major impact on the audience. We also suggest new directions for future interdisciplinary research.

This paper addresses a safe path planning problem for UAV urban navigation, under uncertain GNSS availability. The problem can be modeled as a POMDP and solved with sampling-based algorithms. However, such a complex domain suffers from high computational cost and achieves poor results under real-time constraints. Recent research seeks to integrate offline learning in order to efficiently guide online planning. Inspired by the state-of-the-art CAMP (Context-specific Abstract Markov decision Process) formalization, this paper proposes an offline process which learns the path constraint to impose during online POMDP solving in order to reduce the policy search space. More precisely, the offline learnt constraint selector returns the best path constraint according to the GNSS availability probability in the environment. Conclusions of experiments, carried out for three environments, show that using the proposed approach allows to improve the quality of a solution reached by an online planner, within a fixed decision-making timeframe, particularly when GNSS availability probability is low.

Dense and complex air traffic scenarios require higher levels of automation than those exhibited by tactical conflict detection and resolution (CD&R) tools that air traffic controllers (ATCO) use today. However, the air traffic control (ATC) domain, being safety critical, requires AI systems to which operators are comfortable to relinquishing control, guaranteeing operational integrity and automation adoption. Two major factors towards this goal are quality of solutions, and transparency in decision making. This paper proposes using a graph convolutional reinforcement learning method operating in a multiagent setting where each agent (flight) performs a CD&R task, jointly with other agents. We show that this method can provide high-quality solutions with respect to stakeholders interests (air traffic controllers and airspace users), addressing operational transparency issues.

The SARS-CoV-2 pandemic has galvanized the interest of the scientific community toward methodologies apt at predicting the trend of the epidemiological curve, namely, the daily number of infected individuals in the population. One of the critical issues, is providing reliable predictions based on interventions enacted by policy-makers, which is of crucial relevance to assess their effectiveness. In this paper, we provide a novel data-driven application incorporating sub-symbolic knowledge to forecast the spreading of an epidemic depending on a set of interventions. More specifically, we focus on the embedding of classical epidemiological approaches, i.e., compartmental models, into Deep Learning models, to enhance the learning process and provide higher predictive accuracy.

Nowadays, the Earth observation systems involve multiple satellites, multiple ground stations, and multiple end-users that formulate various observation requests. These requests might be heterogeneous (stereoscopic observations, periodic observations, systematic observations, etc.), and one difficulty is that the search space defined by the possible ways of performing the requests given the multiple satellites and ground stations available is huge. This paper studies several combinatorial optimization techniques for solving such an operational problem, including a constraint programming approach and parallel scheduling techniques that take advantage of the problem structure. These algorithms are evaluated on realistic instances involving various request types and objective functions depending on the cloud cover conditions, that highly impact the quality of the images collected.

In recent years, pervasive digitalization has affected the industrial world, including the oil and gas sector. With more and more data becoming available, Machine Learning algorithms have become a promising tool to improve Predictive Maintenance operations. In this work, we have designed an alerting system that notifies the site operator with an adequate advance when an anomaly is going to occur. In particular, we focus our analysis on the stabilization column of an Oil Stabilization Facility to prevent the column bottom temperature to overcome safety boundaries. The experimental analysis demonstrates that our system provides reliable results, in terms of both identified anomalies and false alarms. In addition, the system is currently under deployment on the company computing infrastructure and the first working version will be available by the end of May 2022.

Categories are important elements of databases of Product Listings, for e-commerce platforms, or of Points of Interest (POIs), for location-based services. However, category annotations are often incomplete, which calls for automatic completion. Hierarchical classification has been proposed as a solution to impute missing annotations. We address this task in one of Naver’s production databases (POIs), in order to enhance its quality. In real-life applications, like ours, however, it is unrealistic to count on the existence of a perfectly annotated training set, and noisy training labels prevent us from casting the task as a straightforward classification problem. In order to overcome this difficulty, we propose an approach that takes into account the type of noise in the training set. We identified that the main deficiency is that the training labels tend to be under-specified i.e. they point to categories found at higher levels of the hierarchy than the correct ones. This results in a lot of under-represented and a few over-represented categories. We call categories that are over-represented, due to under-specified labels, joker classes. To allow robust learning in the presence of joker classes we propose a simple and effective approach: First, we detect problematic categories, i.e. joker classes, based on the misclassifications of an initial hierarchical classifier. Then we re-train from scratch, introducing a weight to the standard cross-entropy loss function that targets incorrect predictions related to joker classes. Our model has enabled the correction of thousands of POIs in our production database.

Given the diffusion of Artificial Intelligence (AI) in numerous domains, experts and practitioners are faced with the challenge of finding the optimal hardware (HW) resources and configuration (hardware dimensioning) under different constraints and objectives (e.g., budget, time, solution quality). To tackle this challenge, we propose an automated tool for HArdware Dimensioning of (AI) Algorithms (HADA), an approach relying on the integration of Machine Learning (ML) models together into an optimization problem, where experts domain knowledge can be injected as well. The ML models encapsulate the data-driven knowledge about the relationships between HW requirements and AI algorithm performances. We show how HADA can be employed to find the best HW configuration that respects user-defined constraints in three different domains.

Flat roofs have become popular also in central and northern Europe during the last decades. One advantage when compared to pitched roofs is that flat roofs are typically significantly cheaper. Furthermore, the roof space can be used also as a garden, a terrace or simply to quite easily install photo-voltaic systems on it. However, flat roofs are known to be prone to drainage and leakage issues. Roof utilization as a garden or the shadowing of installed photo-voltaic systems magnify this problem. For these reasons, installing moisture sensors inside the roof in order to monitor the moisture levels is one possibility to detect roof damages early and keep repairing costs low. In this paper we report on first results of an industrial project that aims to go one step further. Based on past sensor values the goal is to predict how moisture levels will progress in the near future and thus be able to identify problems before they become critical.

Mild Cognitive Impairment (MCI) brings an increased risk of progressing to Alzheimer’s disease (AD). Early identification of a risk of MCI progression could help patients get early treatment to slow progress of the disease. We used 3D Stereotactic Surface Projections (SSP) of Positron Emission Tomography (PET) brain images to train a classification model to identify MCI patients at risk of progressing to AD, which achieved 88.0% accuracy, 85.3% sensitivity, and 90.6% specificity. For model transparency purposes, we generated saliency map explanations from the trained model and evaluated these using radiologist feedback.