Search

In this paper we present ELeaRNT an evolutionary strategy which evolves rich neural network topologies in order to find an optimal domain specific non linear function approximator with a good generalization performance. The neural networks evolved by the algorithm have a feed forward topology with shortcut connections and arbitrary activation functions at each layer. This kind of topologies has not been thoroughly investigated in literature, but is particularly well suited for non linear regression tasks. The experimental results prove that, in such tasks, our algorithm can build, in a completely automated way, neural network topologies able to outperform classic neural network models designed by hand. Also when applied to classification problems, the performance of the obtained neural networks is fully comparable to that of classic neural networks and in some cases noticeably better.

This paper focuses on on-line anomaly detection in video traffic surveillance systems. Markov chain (MC) have been proposed already in computer and network intrusion detection. We applied them to the traffic domain and we propose to extend the classical MC (modeling all the behaviors in the scene) with an approach that evaluates in parallel a set of behavior specific MC. Such separate MCs are more discriminatory than a single MC for all the behaviors, allowing our approach to detect anomalies resulting from joining segments of normal behaviors. The learning of such models is done by using sequences of labeled normal behaviors and discretizing the image plane by using a simple grid. The approach has been validated on traffic surveillance videos, and experimental results show good performance both in terms of precision and recall.

In this paper we propose an approach to color classification and image segmentation in non-stationary environments. Our goal is to cope with changing illumination condition by on-line adapting both the parametric color model and its structure/complexity. Other authors used parametric statistics to model color distribution in segmentation and tracking problems, but with a fixed complexity model. Our approach is able to on-line adapt also the complexity of the model, to cope with large variations in the scene illumination and color temperature.

Patients age has been estimated in healthy population by means of the heart rate variability (HRV) parameters to assess the potentiality of HRV indexes as a biomarker of age. A long-term analysis of HRV has been performed, computing linear time and frequency domain parameters as well as non-linear metrics, in a dataset of 113 healthy subjects (age range 20-85 years old). The principal component analysis has been used to capture age-related influence on HRV and then three different models have been applied to predict subjects age: a robust linear regressor (RLR), a feedforward neural network (FFNN) and a radial basis function neural network (RBFNN). A good prediction of patient age has been obtained (using all principal components, the Pearson correlation coefficient between predicted and real age: RLR=0.793; FFNN=0.872; RBFNN=0.829), even if an overestimation in younger subjects and an underestimation in older ones may be observed. The important and complementary contribution of non-linear indexes to aging related HRV modifications has also been underlined.

In this paper we present a genetic approach that uses a Bayesian fitness function to the design of rich neural network topologies in order to find an optimal domain-specific non-linear function approximator with good generalization performance. Rich neural networks have a feed-forward topology with shortcut connections and arbitrary activation functions at each layer. This kind of topologies is particularly well suited for non-linear regression tasks, but it may suffer for overfilling issues. In this paper we present a Bayesian fitness function to effectively apply genetic algorithms with these models obtaining, in a completely automated way, models well-matched to the problem, with good generalization capability, and low complexity.

Real-time monitoring of the laser-based applications is becoming a main issue for quality analysis in the steel manufacturing industry. The paper suggests a solution achieving an automated real-time quality inspection in laser welding applications. A composite system composed of soft computing and traditional techniques has been considered for its positive impact on the reduced computation compared with more traditional approaches.

In order to support the residual communication capabilities of verbal impaired peoples softwares allowing Augmentative and Alternative Communication (AAC) have been developed. AAC communication software aids provide verbal disables with an electronic table of AAC languages (i.e. Bliss, PCS, PIC, etc.) symbols in order to compose messages, exchange them via email, or vocally synthetize them, and so on. A current open issue, in thins kind of software, regards human-computer interaction in verbal impaired people suffering motor disorders. They can adopt only ad-hoc input device, such as buttons or switches, which require an intelligent automatic scansion of the AAC symbols table in order to compose messages. In such perspective we have developed Caba2l an innovative composition assistant exploiting an user linguistic behavior model adopting a semantic/probabilistic approach for predictive Bliss symbols scansion. Caba2l is based on an original discrete implementation of auto-regressive hidden Markov model called DAR-HMM and it is able to predict a list of symbols as the most probable ones according to both the previous selected symbol and the semantic categories associated to the symbols. We have implemented the composition assistant as a component of Bliss2003 an AAC communication software centered on Bliss language and experimentally validated it with both synthetic and real data.

We present the framework we have adopted to implement robust sensing in the Milan Robocup F-2000 Team. The main issue concerns the definition of symbolic models both for the single agent and for the whole multi-agent system. Information about physical objects is collected by intelligent sensors and anchored to symbolic concepts. These are used both to control the robots through a behavior-based system, and to improve the global knowledge about the environment.

We present Fun2maS, the Milan Robocup Team. In its implementation we have faced many aspects: hardware (electronics, mechanics), sensors (omnidirectional vision), behaviors (fuzzy behaviors management and composition), multi-agent coordination (strategy and tactics), and adaptation of the team behavior. We could fully exploit the characteristics of all the components thanks to the modular design approach we have adopted. All the modules have also been designed to be used in generic applications, and we have already adopted most of them for surveillance, mapping, guidance, and document delivery tasks. © 2002 Springer-Verlag Berlin Heidelberg.

The success of any reinforcement learning (RL) application is in large part due to the design of an appropriate reinforcement function. A methodological framework to support the design of reinforcement functions has not been defined yet, and this critical and often underestimated activity is left to the ability of the RL application designer. We propose an approach to support reinforcement function design in RL applications concerning learning behaviors for autonomous agents. We define some dimensions along which we can describe reinforcement functions; we consider the distribution of reinforcement values, their coherence and their matching with the designer's perspective. We give hints to define measures that objectively describe the reinforcement function; we discuss the trade-offs that should be considered to improve learning and we introduce the dimensions along which this improvement can be expected. The approach we are presenting is general enough to be adopted in a large number of RL projects. We show how to apply it in the design of learning classifier systems (LCS) applications. We consider a simple, but quite complete case study in evolutionary robotics, and we discuss reinforcement function design issues in this sample context.

Anchoring is the activity an agent does to classify the objects it perceives and to track them. With the increasing abilities of mobile robots, anchoring is becoming a major research issue to make robots perform their task in unstructured and partially unknown environments. We propose a model to represent knowledge in an agent, possibly participating to a multi-agent system, showing that the anchoring problem can be successfully faced by well known AI techniques. We discuss how this model can reliably support the instantiation of concepts by aggregating percepts affected by uncertainty, sensed by several sensors, and obtained from different agents.

We discuss some issues concerning the application of learning classifier systems to real-valued applications. In particular, we focus on the possibility of classifying data by crisp and fuzzy intervals, showing the effect of their granularity on the learning performance. We introduce the concept of sensorial cluster and we discuss the difference between cluster aliasing and perceptual aliasing. We show the impact of different choices on the performance of both crisp and fuzzy learning classifier systems applied to a mobile, autonomous, robotic agent.

An alternative DSS which models the behaviour of the Heart Rate Variability (HRV) signal linked to stable (NREM) and instable (REM) cerebral waves during sleep and a probabilistic model of the sleep stages transitions for decision was developed. Time-Varying Autoregressive Models (TVAMs) were used as feature extractor while Hidden Markov Models (HMM) was used as time series classifier. 24 full polysomnography recordings from healthy sleepers were used for the analysis and those were separated in two sets of

Persons suffering from motor disorders have limited possibilities for communicating and normally require assistive technologies to fulfill this primary need. Promising means of providing basic communication abilities to subjects affected by severe motor impairments include brain-computer interfaces (BCIs), that is, systems that directly translate brain signals into device commands, bypassing any muscle or nerve mediation. To date, the use of BCIs for effective verbal communication is yet an open issue, primarily due to the low rates of information transfer that can be achieved with this technology. Still, performance of BCI spelling applications could be considerably improved by a smart user interface design and by the adoption of natural language processing (NLP) techniques for text prediction. The objective of this work is to suggest an approach and a user interface for BCI spelling applications combining state-of-the-art BCI and NLP techniques to maximize the overall communication rate of the system. The BCI paradigm adopted is motor imagery, that is, when the subject imagines moving a certain part of the body, he/she produces modifications to specific brain rhythms that are detected in real-time through an electroencephalogram and translated into commands for a spelling application. By maximizing the overall communication rate, our approach is twofold: on one hand, we maximize the information transfer rate from the control signal, on the other hand, we optimize the way this information is employed for the purpose of verbal communication. The achieved results are satisfactory and comparable with the latest works reported in literature on motor-imagery BCI spellers. For the three subjects tested, we obtained a spelling rate of respectively 3 char/min, 2.7 char/min, and 2 char/min.

Coordination, planning, and reactivity are important for successful teams of autonomous robots, in dynamic adversarial domains. In this paper, we propose a fuzzy cognitive model to integrate coordination, planning and reactive behaviors in a team of cooperating robots. In our architecture, behavioral modules are used as high-level macro-actions that compose structured plans defined by a flexible multi-agent coordination system. The use of an unifying cognitive model provides an effective tool for seamless integration of reactive and deliberative components in the robots, gaining as much as possible from the presence of different skills and competencies in the team. The control model is designed to be tuned and adapted on-line so that the team strategies and the role of robots in the control schemata can be automatically modified to face different opponent teams, and changes in robot capabilities.

In this work we describe a simulation environment for an autonomous all-terrain mobile robot. To allow for extensive test and verification of the high-level perception, planning, and trajectory control modules, the low-level control systems, the sensors, and the vehicle dynamics have been modeled and simulated by means of the V-Rep 3D simulator. We discuss the overall, i.e., high and low-level, software architecture and we present some validation experiments in which the behavior of the real system is compared with the corresponding simulations.

Stuck-pipe phenomena can have disastrous effects on drilling performance, with outcomes that may range from time delays to loss of expensive machinery. In this work, we develop three indicators based on the mudlog data, which aim to detect three different physical phenomena associated to the insurgence of a sticking. In particular, two indices target respectively the detection of translational and rotational motion issues, while the third index concerns the wellbore pressure. A statistical model that relates these features with the documented stuck-pipe events is then developed using machine learning. The resulting model takes the form of a depth-based map of the risk of incurring into a stuck-pipe, updated in real time. Preliminary experimental results on the available dataset indicate that the use of the proposed model and indicators can help mitigate the stuck-pipe issue.

In this study, surface ECG signals recoded during atrial fibrillation (AF) episodes have been investigated to detect signs of spontaneous termination and to derive an automatic classifier of terminating (T) and nonterminating (N) AF events. The ECG signals consisted in Holier recordings coming from the 2004 Computers in Cardiology Challenge database, A set of features have been extracted from the ECG signals and the related RR interval series including both linear and nonlinear indexes. In the training dataset, we observed a prolonged dominant atrial cycle length (DACL) passing from N to T accompanied by an increased of residual ECG (rECG) power. Concerning the RR interval variability a reduction of mean RR interval and Regularity (R) and an increase of approximated entropy (ApEn) have been documented. These features were used to train a feed-forward neural network which was employed for the automatic classification of the challenge test set. Score of the classifier was encouraging: 26/30 episodes were correctly classified.

This study analyses the spectral components of the heart rate fluctuations of a new contact-less technology for sleep evaluation. Both heart beat interval (HBI) and movement activity were extracted from the multichannel ballistocardiographic (BCG) measurements, based on Emfit sensor foils placed into bed mattress. Powers spectral densities (PSD) of HBI have been compared with the ones obtained from the standard ECG during sleep stage 2. In addition, spectral features obtained from the contact-less technology and standard ECG has been used to automatically classify the sleep macrostructure through a time-varying autoregressive model and a Hidden Markov Model. Whole night recordings from six subjects were analyzed in this study. Spectral components did not show significant differences between the two measurements. Further, contactless technology achieved a total accuracy of 83 % and kappa index of 0.42, while standard ECG achieved an accuracy of 84 % and kappa index of 0.43 when compared to clinical sleep staging from polysomnography.

In this paper, we present a dead reckoning sensor to support reliable odometry on mobile robots. This sensor is based on a pair of optical mice rigidly connected to the robot body and its main advantages are 1) this localization system is independent from the kinematics of the robot, 2) the measurement given by the mice is not subject to slipping, since they are independent from the traction wheels, nor to crawling, since they measure displacements in any direction 3) it is a low-cost solution with a precision comparable to classical shaft encoders. We present the mathematical model of the sensor, its implementation, and some experimental evaluations using the standard UMBmark benchmark for odometry.

The adoption of Artificial intelligence (AI) technologies is steadily increasing. However, to become fully pervasive, AI needs resources at the edge of the network. The cloud can provide the processing power needed for big data, but edge computing is close to where data are produced and therefore crucial to their timely, flexible, and secure management. In this paper, we introduce the AI-SPRINT project, which will provide solutions to seamlessly design, partition, and run AI applications in computing continuum environments. AI-SPRINT will offer novel tools for AI applications development, secure execution, easy deployment, as well as runtime management and optimization: AI-SPRINT design tools will allow trading-off application performance (in terms of end-to-end latency or throughput), energy efficiency, and AI models accuracy while providing security and privacy guarantees. The runtime environment will support live data protection, architecture enhancement, agile delivery, runtime optimization, and continuous adaptation.

In this paper, we propose a modeling paradigm that uses fuzzy sets to represent concepts on which control modules of a behavior-based autonomous robot operate. The primitives defined in the modeling paradigm are expressive enough to represent the knowledge needed by planning, coordination, and reactive control of a multi-robot control system. At the same time, it provides a well-founded tool to represent in a compact way the data interpretations needed to reason effectively about what is happening in the world and what is desired to happen. This modeling paradigm makes the design of behavior, planning, and coordination modules easy, since its primitives are simple and expressive. Moreover, it provides a sound framework to deal with uncertainty in sensing and world modeling.

Lateral control is one of the primary requirements of an autonomous vehicle. This task is generally performed using complex pipelines, which include line detection trough neural network processing, vehicle state estimation, and planning. What we propose in this paper is an alternative end-to-end approach to the problem. Images acquired by a camera mounted on the vehicle are processed by two convolutional neural networks to directly retrieve the steering command. In particular, we propose an architecture built using two connected neural networks, one to predict the scenario the vehicle is facing and one, conditioned on possible situations, to predict the steering command. In our work, we also analyze the potential of a computer-generated dataset for a demanding task like end-to-end learning, where the image quality is fundamental. All the training is then performed on synthetic images, while the testing is done on real data acquired by an experimental vehicle.

This study proposes an alternative evaluation of obstructive sleep apnea (OSA) based on ECG signal during sleep time. OSA is a common sleep disorder produced by repetitive occlusions in the upper airways. This respiratory disturbance produces a specific pattern on ECG. Extraction of ECG characteristics, as heart rate variability (HRV) and peak R area, offers alternative measures for a sleep apnea pre-diagnosis. 50 recordings coming from the apnea Physionet database were used in the analysis, this database is part of the 70 recordings used for the Computer in Cardiology challenge celebrated in 2000. A bivariate autoregressive model was used to evaluate beat-by-beat power spectral density of HRV and R peak area. K-nearest neighbor (KNN) supervised learning classifier was employed for categorizing apnea events from normal ones, on a minute-by-minute basis for each recording. Data were split into two sets, training and testing set, each one with 25 recordings. The classification results showed an accuracy higher than 85% in both training and testing. In addition it was possible to separate completely between apnea and normal subjects and almost completely among apnea, normal and borderline subjects.

In this paper we describe Bliss2003, an Information and Communication Technology (ICT) aid for verbal impaired people supporting the use of Augmentative and Alternative Communication (AAC) languages. Bliss2003 allows to compose messages in Bliss and other AAC languages (i.e. PCS, PIC, etc.), to translate them in natural language, to send and receive them via email, or to vocally synthetize them. Bliss2003 is characterized by a predictive module that allows for a more efficient selection of graphical symbols and more natural sessions of communication by adapting a model of the user language behavior.

Data from benchmarkruns of a simulated ROS2 autonomous navigation system with different combinations of robot model, odometry sensor, environment (always static), local planner (TEB, DWB, RPP), and global planner (SMAC, NavFN). The results multiple performance metrics related to localiaztion and local planning. The results are generated with the code inhttps://github.com/AIRLab-POLIMI/local_planning_performance_modelling/tree/foxy-devel.

This deliverable, which is the first release of the market analysis and forecast report for AI-SPRINT, provides a detailed assessment of the market status-quo and market forecast for edge computing with a specific focus on AI and other enabling technologies. It also gives a deep dive into the market forecast of the three main use-cases of the AI-SPRINT project. The European enterprise edge market has grown significantly in the last couple of years, with new solutions developed to address specific industries and use cases as more organisations have started to understand the benefits and opportunities behind edge and other emerging technologies. The COVID-19 pandemic has reshuffled the investment plans of many organisations in the world, in some cases stimulating the need for edge to ensure business continuity, to improve the decision-making process and to support remote operations. European enterprise edge spending is expected to reach €39.6 billion by 2025, driven by the tremendous role of edge in enabling low latency use cases, deployments in locations that are either remote or without IT support staff, and supporting new vendor propositions and a supply chain that is being built up. Major findings in this document include the following: European enterprise edge spending will grow from €18 billion in 2020 to €39.6 billion in 2025, with a five-year Compound Annual Growth Rate (CAGR) of 18%, outpacing the total ICT market average that is forecast to grow at single digits over the same period. Compared with Central and Eastern Europe (CEE), Western Europe sees a more established and developed edge scenario. Nevertheless, edge adoption in CEE is expected to increase over the forecast period, especially within key verticals. Within the three main technology groups, services (including professional and provisioned services) will continue to make half of the total market spending. Hardware will outpace the market average, reaching a fourth of the market, as new hybrid edge infrastructures drive innovations at the edge. Edge supports a wide array of other emerging technology markets referred to as domains, that are driving the overall edge market and generating value outside core data centers. These are Internet of Things (IoT) as the largest domain within the European edge market, followed by artificial intelligence (AI) with tremendous capabilities on real-time data processing, augmented reality (AR)/virtual reality (VR), robotics and drones. The distribution and services sector along with manufacturing and resources have the largest share of European enterprise edge spending. Despite the pandemic heavily hitting some industries within those sectors, edge use cases linked to the Internet of Things (IoT) and AI domains will drive developments, especially in edge environments where edge bridges the gap between IT and operations technologies (OT). The report double clicks into the market dimension for edge solutions by type of edge technology, and by industry, looking also at the AI domain in the context of AI-SPRINT. In particular, the report provides a detailed analysis on the three use cases in the AI- SPRINT context: maintenance and inspection, personalised healthcare and farming 4.0, looking also at the drivers and the barriers for the edge market overall and in the context of specific industry use cases. In the market context chapter, the report presents the significant market developments for the edge market but also an analysis of how the ecosystem is a pivotal point for deploying edge solutions, particularly in the context of the three AI-SPRINT cases., {"references":["Project deliverable","Open Access"]}

The aim of the AI-SPRINT “Artificial intelligence in Secure PRIvacy-preserving computing coNTinuum” project is to develop a platform composed of design and runtime management tools to seamlessly design, partition and operate Artificial Intelligence (AI) applications among the current plethora of cloud-based solutions and AI-based sensor devices (i.e., devices with intelligence and data processing capabilities), providing resource efficiency, performance, data privacy, and security guarantees. The aim of this deliverable is to review the state-of-the-art in techniques used in the development and operation of AI applications in computing continua and the related technologies. The deliverable starts with an overview of AI applications and edge computing market trends. It examines the standardisation and opensource landscape, where many of the emerging standards are mainly focused on the network layer, also highlighting opportunities for open-source community involvement to expand technology availability. Then, the deliverable provides a background on AI applications design, also considering some advanced design trends (e.g., Network Architecture Search, Federated Learning, Deep Neural Networks partitioning) introducing the software development frameworks and hardware solutions which allow to run such applications in a computing continuum. We then report an overview of the state-of-the-art of the solutions offered by the main cloud market players and the European providers. We also extensively discuss existing solutions for applications deployment, monitoring, runtime management, and scheduling considering the emerging Function as a Service paradigm. In the last part of the deliverable, we report an overview of the performance modelling solutions, security, and privacy problems for AI applications in edge environments. It is found that some areas, such as component placement and design space exploration with privacy preservation and performance guarantees, are fairly under-developed, since solutions tailored for AI applications are lacking, and thus offer an opportunity for innovation., {"references":["Project deliverable","Open Access"]}

This document, developed by the AI-SPRINT project, represents an initial version of the platform Architecture deliverable (the final due at M24) that offers an overview of the architectural design. The focus of this document is the identification of different components and interfaces within the AI-SPRINT architectural building blocks, as well as the different strategies necessary for a successful deployment. The role of each of the blocks is defined, as well as the required interfaces for the communication between blocks. The architectural choices are built taking into account the different types of requirements imposed by the applications and services running in the framework through the analysis of the three use cases specified in the project proposal and detailed in AI-SPRINT Deliverable D1.2 - Requirements Analysis., {"references":["Project deliverable","Open Access"]}

D1.2 summarizes the initial requirements analysis of the AI-SPRINT project. The outcome of this analysis is a list of requirements for all tools which will be evolved or developed within the project. This analysis will guide all the AI-SPRINT technical activities. The aim of AI-SPRINT (Artificial Intelligence in Secure PRIvacy-preserving computing coNTinuum) is to define a novel framework for developing and operating AI applications, together with their data, exploiting computing continuum environments, which include resources from the edge up to the cloud. AI-SPRINT will offer novel tools for AI applications development, secure execution, easy deployment, as well as runtime management and optimization. AI-SPRINT tools will allow trading-off application performance (in terms of end-to-end latency or throughput), energy efficiency, and AI models accuracy while providing security and privacy guarantees. AI-SPRINT framework will support AI applications data protection, architecture enhancement, agile delivery, runtime optimization, and continuous adaptation. A critical success factor in the development of high quality, state of the art, software frameworks is a deep understanding of the requirements of users of the framework (e.g. developers, AI experts, service providers etc.) and of end-users of the application and services created using the framework (e.g. farmers, stroke patience, maintenance engineers etc.). The elicitation of the AI-SPRINT requirements started with the description of Personas, i.e. fictional images of typical AI-SPRINT users with their specific interests, experiences and demands. User Stories and Requirements were then formulated by looking at the AI-SPRINT from the perspective of the selected Personas. The providers of the three real-life Use Cases wrote the first draft of the requirements, the tool providers reviewed and enhanced these requirements from a more technical point of view, also keeping in mind current state of the art solutions (see deliverable D1.1). The requirements analysis helped to outline the importance of some critical AI-SPRINT features. The ability to develop cloud/edge agnostic applications and to define at runtime which component will run on the edge, on the cloud or on a hybrid-cloud setup will be a pivotal strength of the tools. The excellent security features of the framework will allow its usage also in security and privacy critical domains (e.g., Personalized Healthcare). The federated training of AI models will also prove to be extremely important in areas in which data cannot be easily shared among the different parties involved in the project. The requirements analysis also showed that the proposed reference architecture (see deliverable D1.3) can be used for all three Use Cases, with minor specific modifications. The proposed project assets (SCAR, PyCOMPs, Krake, SCONE etc.) present a very good starting point for the development of the framework; the tools need to be adapted and especially integrated to form a coherent, easy to use framework. It must be mentioned that the requirements analysis process is intended to be a continuous process, hence this report does not coincide with the end of requirement analysis. A requirement repository was defined and will be continuously updated until month 24., {"references":["Project deliverable","Open Access"]}