Deep reinforcement learning has made significant development in recent years, and it is currently applied not only in simulators and games but also in embedded systems. However, when implemented in a real-world context, reinforcement learning is frequently shown to be unstable and incapable of adapting to realistic situations, particularly when directing a large number of agents. In this paper, we develop a decentralized architecture for reinforcement learning to allow multiple agents to learn optimal control policies on their own devices of the same kind but in varied environments. For such multiple agents, the traditional centralized learning algorithm usually requires a costly or time-consuming effort to develop the best-regulating policy and is incapable of scaling to a large-scale system. To address this issue, we propose a decentralized reinforcement learning algorithm (DecRL) and information exchange scheme for each individual device, in which each agent shares the individual learning experience and information with other agents based on local model training. We incorporate the algorithm into each agent in the proposed collaborative architecture and validate it in the telecommunication domain under emergency conditions, in which a macro base station (BS) is broken due to a natural disaster, and three unmanned aerial vehicles carrying BSs (UAV-BSs) are deployed to provide temporary coverage for missioncritical (MC) users in the disaster area. Based on the findings, we show that the proposed decentralized reinforcement learning algorithm can successfully support multi-agent learning, while the learning speed and service quality can be further enhanced.

Deep reinforcement learning has advanced signifi-cantly in recent years, and it is now used in embedded systems in addition to simulators and games. Reinforcement Learning (RL) algorithms are currently being used to enhance device operation so that they can learn on their own and offer clients better services. It has recently been studied in a variety of industrial applications. However, reinforcement learning, especially when controlling a large number of agents in an industrial environment, has been demonstrated to be unstable and unable to adapt to realistic situations when used in a real-world setting. To address this problem, the goal of this study is to enable multiple reinforcement learning agents to independently learn control policies on their own in dynamic industrial contexts. In order to solve the problem, we propose a dynamic multi-agent reinforcement learning (dynamic multi-RL) method along with adaptive exploration (AE) and vector-based action selection (VAS) techniques for accelerating model convergence and adapting to a complex industrial environment. The proposed algorithm is tested for validation in emergency situations within the telecommunications industry. In such circumstances, three unmanned aerial vehicles (UAV-BSs) are used to provide temporary coverage to mission-critical (MC) customers in disaster zones when the original serving base station (BS) is destroyed by natural disasters. The algorithm directs the participating agents automatically to enhance service quality. Our findings demonstrate that the proposed dynamic multi-RL algorithm can proficiently manage the learning of multiple agents and adjust to dynamic industrial environments. Additionally, it enhances learning speed and improves the quality of service.

Continuous experimentation (CE) refers to a set of practices used by software companies to rapidly assess the usage, value, and performance of deployed software using data collected from customers and systems in the field using an experimental methodology. However, despite its increasing popularity in developing web-facing applications, CE has not been studied in the development process of business-to-business (B2B) mission-critical systems. By observing the CE practices of different teams, with a case study methodology inside Ericsson, we were able to identify the different practices and techniques used in B2B mission-critical systems and a description and classification of the four possible types of experiments. We present and analyze each of the four types of experiments with examples in the context of the mission-critical long-term evolution (4G) product. These examples show the general experimentation process followed by the teams and the use of the different CE practices and techniques. Based on these examples and the empirical data, we derived the HURRIER process to deliver high-quality solutions that the customers value. Finally, we discuss the challenges, opportunities, and lessons learned from applying CE and the HURRIER process in B2B mission-critical systems. 

Continuous deployment has been practiced for many years by companies developing web- and cloud-based applications. To succeed with continuous deployment, these companies have a strong collaboration culture between the operations and development teams. In addition, these companies use AI, analytics, and big data to assist with time-consuming postdeployment activities such as continuous monitoring and fault identification. Thus, the term AIOps has evolved to highlight the importance and difficulty of maintaining highly available applications in a complex and dynamic environment. In contrast, software-intensive embedded systems often provide customer product-related services, such as maintenance, optimization, and support. These services are critical for these companies as they provide significant revenue and increase customer satisfaction. Therefore, the objective of our study is to gain an in-depth understanding of the impact of continuous deployment on product-related services provided by software-intensive embedded systems companies. In addition, we aim to understand how AIOps can support continuous deployment in the context of software-intensive embedded systems. To address this objective, we conducted a case study at a large and multinational telecommunications systems provider focusing on the radio access network (RAN) systems for 4G and 5G networks. The company provides RAN products and three complementing services: rollout, optimization, and customer support. The results from the case study show that the boundaries between product-related services become blurry with continuous deployment. In addition, product-related services, which were conducted in sequence by independent projects, converge with continuous deployment and become part of the same project. Further, AIOps platforms play an important role in reducing costs and increasing postdeployment activities' efficiency and speed. These results show that continuous deployment has a profound impact on the software-intensive system's provider service organization. The service organization becomes the connection between the R&D organization and the customer. In order to cope with the increased speed of releases, deployment and postdeployment activities need to be largely automated. AIOps platforms are seen as a critical enabler in managing the increasing complexity without increasing human involvement.

Artificial intelligence (AI) and the use of machine learning (ML) and deep learning (DL) technologies are becoming increasingly popular in companies. These technologies enable companies to leverage big quantities of data to improve system performance and accelerate business development. However, despite the appeal of ML/DL, there is a lack of systematic and structured methods and processes to help data scientists and other company roles and functions to develop, deploy and evolve models. In this paper, based on multi-case study research in six companies, we explore practices and challenges practitioners experience in developing ML/DL models as part of large software-intensive embedded systems. Based on our empirical findings, we derive a conceptual framework in which we identify three high-level activities that companies perform in parallel with the development, deployment and evolution of models. Within this framework, we outline activities, iterations and triggers that optimize model design as well as roles and company functions. In this way, we provide practitioners with a blueprint for effectively integrating ML/DL model development into the business to achieve better results than other (algorithmic) approaches. In addition, we show how this framework helps companies solve the challenges we have identified and discuss checkpoints for terminating the business case.

Artificial Intelligence (AI) engineering is an engineering discipline that is concerned with all aspects of development and evolution of AI systems (that is, systems that include AI components). AI engineering is primarily an extension of software engineering, but it also includes methods and technologies from data science and AI in general.

Fast and reliable connectivity is essential to enhance situational awareness and operational efficiency for public safety mission-critical (MC) users. In emergency or disaster circumstances, where existing cellular network coverage and capacity may not be available to meet MC communication demands, deployable-network-based solutions such as cells-on-wheels/wings can be utilized swiftly to ensure reliable connection for MC users. In this paper, we consider a scenario where a macro base station (BS) is destroyed due to a natural disaster and an unmanned aerial vehicle carrying BS (UAV-BS) is set up to provide temporary coverage for users in the disaster area. The UAV-BS is integrated into the mobile network using the 5G integrated access and backhaul (IAB) technology. We propose a framework and signalling procedure for applying machine learning to this use case. A deep reinforcement learning algorithm is designed to jointly optimize the access and backhaul antenna tilt as well as the three-dimensional location of the UAV-BS in order to best serve the on-ground MC users while maintaining a good backhaul connection. Our result shows that the proposed algorithm can autonomously navigate and configure the UAV-BS to improve the throughput and reduce the drop rate of MC users.

In recent years, the application of artificial intelligence (AI) has become an integral part of a wide range of areas, including software engineering. By analyzing various data sources generated in software engineering, it can provide valuable insights into customer behavior, product performance, bugs and errors, and many more. In practice, however, AI for software analytics and business intelligence often remains at a prototypical stage, and the results are rarely used to make decisions based on data. To understand the underlying causes of this phenomenon, we conduct an explanatory case study consisting of and interview study and a survey on the challenges of realizing and utilizing artificial intelligence in the context of software-intensive businesses. As a result, we identify a vicious circle that prevents practitioners from moving from prototypical AI-based analytics to continuous and productively usable software analytics and business intelligence solutions. In order to break the vicious circle in a targeted manner, we identify a set of solutions based on existing literature as well as the previously conducted interviews and survey. Finally, these solutions are validated by a focus group of experts. (C) 2021 Elsevier Inc. All rights reserved.

AI in the Engineering, Procurement and Construction (EPC) industry has not yet a proven track record in large-scale projects. Since AI solutions for industrial applications became available only recently, deployment experience and lessons learned are still to be built up. Several research papers exist describing the potential of AI, and many surveys and white papers have been published indicating the challenges of AI deployment in the EPC industry. However, there is a recognizable shortage of in-depth studies of deployment experience in academic literature, particularly those focusing on the experiences of EPC companies involved in large-scale project execution with high safety standards, such as the petrochemical or energy sector. The novelty of this research is that we explore in detail the challenges and obstacles faced in developing and deploying AI in a large-scale project in the EPC industry based on real-life use cases performed in an EPC company. Those identified challenges are not linked to specific technology or a company's know-how and, therefore, are universal. The findings in this paper aim to provide feedback to academia to reduce the gap between research and practice experience. They also help reveal the hidden stones when implementing AI solutions in the industry.

Continuous deployment has become a widely used practice in web-based software applications. Deploying a new software version to production is a seamless automated process executed thousands of times per day. Continuous deployment reduces the time between a code commit and that commit is active in production. While continuous deployment promises many advantages to software development organizations, the adoption of continuous deployment in the software-intensive embedded systems industry is limited. Several empirical studies have highlighted the challenges associated with software-intensive embedded systems. However, very few studies, if any at all, have attempted to provide a practical approach to realize continuous deployment to these systems. This paper proposes a Controlled Continuous Deployment (CCD) approach, which considers the constraints software-intensive embedded systems have, such as high reliability and availability requirements, limited possibility for rollback after deployment, and the high volume of in-service systems in the market. We derived the approach by conducting a case study at Ericsson AB, focusing on three Radio Access Networks (RAN) technologies embedded software used in 3G, 4G, and 5G mobile networks.  

Supporting customers after they acquire the product is essential for companies producing and selling software-intensive embedded systems products. Generally, customer support is the first interaction point between the product users and the product vendor. Customer support is often engaged with answering customers' questions, troubleshooting, fault identification, and fixing product faults. While continuous deployment advocates for closer cooperation between the ones operating the software and the ones developing it, the means of such collaboration in general and the role of customer support, in particular, has not been addressed in the context of software-intensive embedded systems. Therefore, to better understand the impact that continuous deployment has on customer support and the role customer support should play in this context, we conducted a case study at a multinational company developing and selling telecommunications networks infrastructure. We focused on the 4th and 5th Generation (4G and 5G) Radio Access Networks (RAN) products, which can be considered a high volume product as they cover more than 80% of the world's population. Our study reveals that customer support needs to transition from a transaction-based and passive function triggered by customer support requests, to take an active role characterized by being proactive and preemptive to cope with the shorter operational time of a software version introduced by continuous deployment. In addition, customer support plays an essential role in making the feedback actionable by aggregating and consolidating feedback data to the R&D organization.

Deep learning (DL) based software systems are difficult to develop and maintain in industrial settings due to several challenges. Data management is one of the most prominent challenges which complicates DL in industrial deployments. DL models are data-hungry and require high-quality data. Therefore, the volume, variety, velocity, and quality of data cannot be compromised. This study aims to explore the data management challenges encountered by practitioners developing systems with DL components, identify the potential solutions from the literature and validate the solutions through a multiple case study. We identified 20 data management challenges experienced by DL practitioners through a multiple interpretive case study. Further, we identified 48 articles through a systematic literature review that discuss the solutions for the data management challenges. With the second round of multiple case study, we show that many of these solutions have limitations and are not used in practice due to a combination of four factors: high cost, lack of skill-set and infrastructure, inability to solve the problem completely, and incompatibility with certain DL use cases. Thus, data management for data-intensive DL models in production is complicated. Although the DL technology has achieved very promising results, there is still a significant need for further research in the field of data management to build high-quality datasets and streams that can be used for building production-ready DL systems. Furthermore, we have classified the data management challenges into four categories based on the availability of the solutions.(c) 2022 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Reinforcement learning, particularly deep reinforcement learning, has made remarkable progress in recent years and is now used not only in simulators and games but is also making its way into embedded systems as another software-intensive domain. However, when implemented in a real-world context, reinforcement learning is typically shown to be fragile and incapable of adapting to dynamic environments. In this paper, we provide a novel dynamic reinforcement learning algorithm for adapting to complex industrial situations. We apply and validate our approach using a telecommunications use case. The proposed algorithm can dynamically adjust the position and antenna tilt of a drone-based base station to maintain reliable wireless connectivity for mission-critical users. When compared to traditional reinforcement learning approaches, the dynamic reinforcement learning algorithm improves the overall service performance of a drone-based base station by roughly 20%. Our results demonstrate that the algorithm can quickly evolve and continuously adapt to the complex dynamic industrial environment.

For companies in the embedded systems domain, digitalization and digital technologies allow endless opportunities for new business models and continuous value delivery. While physical products still provide the core revenue, these are rapidly being complemented with offerings that allow for recurring revenue and that are based on software, data and artificial intelligence (AI). However, while new digital offerings allow for fundamentally new and recurring revenue streams and continuous value delivery to customers, the creation of these proves to be a challenging endeavour. In this paper, we study how companies explore ways to create new or additional value with the intention to complement their product portfolio with offerings that allow for recurring revenue. Based on multi-case study research, we identify the key challenges that companies in the embedded systems domain experience and we derive four organizational patterns that we see slow down innovation. Second, we present a framework outlining alternative types of offerings to customers. Third, we provide a value taxonomy in which we detail the different types of offerings and the value these provide to customers. For each value offering, we indicate whether this offering is (1) static or evolving, (2) bundled or unbundled, (3) free or monetized, and we provide examples from the case companies we studied.

With the increased availability of new and better computer processing units (CPUs) as well as graphical processing units (GPUs), the interest in statistical learning and deep learning algorithms for classification tasks has grown exponentially. These classification algorithms often require the presence of fully labeled instances during the training period for maximum classification accuracy. However, in industrial applications, data is commonly not fully labeled, which both reduces the prediction accuracy of the learning algorithms as well as increases the project cost to label the missing instances. The purpose of this paper is to survey the current state-of-the-art literature on machine learning algorithms that are used for assisted or automatic labeling and to understand where these are used. We performed a systematic mapping study and identified 52 primary studies relevant to our research. This paper provides three main contributions. First, we identify the existing machine learning algorithms for labeling and we present a taxonomy of these algorithms. Second, we identify the datasets that are used to evaluate the algorithms and we provide a mapping of the datasets based on the type of data and the application area. Third, we provide a process to support people in industry to optimally label their dataset. The results presented in this paper can be used by both researchers and practitioners aiming to improve the missing labels with the aid of machine algorithms or to select appropriate datasets to compare new state-of-the art algorithms in their respective application area.

Data has been claimed to be the new oil of the 21st century as it has seen to be able both to improve the existing products and services as well as to create new revenue streams for its utilizing company with a secondary customers base. However, while there is active streams of research for developing machine learning and data science methods, considerably less has been done to understand and characterize data business activities in the software-intensive companies. This study uses a multiple case study approach in the software-intensive embedded system domain. Four large international embedded system companies were selected as the case study subjects. The objective is to understand how the case companies are developing their activities for successful utilization of the data. The study identifies six distinct stages with their own challenges. In addition, this study serves as a starting for further work for supporting software-intensive embedded system companies to start data business.

Artificial intelligence is increasingly becoming important to businesses since many companies have realized the benefits of applying Machine Learning (ML) and Deep Learning (DL) into their operations. Nevertheless, ML/DL technologies' industrial development and deployment examples are still rare and generally confined within a small cluster of large international companies who are struggling to apply ML more broadly and deploy their use cases at a large scale. Meanwhile, current AI market has started offering various solutions and services. Thus, organizations must understand how to acquire AI technology based on their business strategy and available resources. This paper discusses the industrial experience of developing and deploying ML/DL use cases to support organizations in their transformation towards AI. We identify how various factors, like cost, schedule, and intellectual property, can be affected by the choice of approach towards ML/DL project development and deployment within large international engineering corporations. As a research result, we present a framework that covers the trade-offs between those various factors and can support engineering companies to choose the best approach based on their long-term business strategies and, therefore, would help to accomplish their ML/DL project deployment successfully.  

Modern release engineering practices such as continuous integration and delivery have allowed software development companies to transition from a long release cycle to a shorter one. The shorter release cycle has led to more software releases available to customers. At the same time, companies developing high-volume software-intensive embedded systems often deliver patch releases and maintenance releases on top of major and minor releases to customers who pick and choose what releases apply to them and decide when to upgrade the system, if to upgrade at all. While release engineering has been studied before in web-based, desktop-based, and embedded software, the focus has been on pre-release activities. Few studies have investigated what happens after the release, particularly the role of tracing software from release to deployment in high-volume software-intensive embedded systems. To address this gap, we conducted a qualitative case study at a multi-national telecommunications systems provider focusing on Radio Access Network (RAN) software. RAN software is a complex and large-scale embedded software used in mobile networks Base Stations (BS), providing software functionality for RAN mobile technologies ranging from 2G to 5G. Our study shed light on post-release software traceability and how it is used in the release engineering process.

In recent years, with more edge devices being put into use, the amount of data that is created, transmitted and stored is increasing exponentially. Moreover, due to the development of machine learning algorithms, modern software-intensive systems are able to take advantage of the data to further improve their service quality. However, it is expensive and inefficient to transmit large amounts of data to a central location for the purpose of training and deploying machine learning models. Data transfer from edge devices across the globe to central locations may also raise privacy and concerns related to local data regulations. As a distributed learning approach, Federated Learning has been introduced to tackle those challenges. Since Federated Learning simply exchanges locally trained machine learning models rather than the entire data set throughout the training process, the method not only protects user data privacy but also improves model training efficiency. In this paper, we have investigated an advanced machine learning algorithm, Deep Neural Decision Forests (DNDF), which unites classification trees with the representation learning functionality from deep convolutional neural networks. In this paper, we propose a novel algorithm, AF-DNDF which extends DNDF with an asynchronous federated aggregation protocol. Based on the local quality of each classification tree, our architecture can select and combine the optimal groups of decision trees from multiple local devices. The introduction of the asynchronous protocol enables the algorithm to be deployed in the industrial context with heterogeneous hardware settings. Our AF-DNDF architecture is validated in an automotive industrial use case focusing on road objects recognition and demonstrated by an empirical experiment with two different data sets. The experimental results show that our AF-DNDF algorithm significantly reduces the communication overhead and accelerates model training speed without sacrificing model classification performance. The algorithm can reach the same classification accuracy as the commonly used centralized machine learning methods but also greatly improve local edge model quality.

A/B experimentation is a known technique for datadriven product development and has demonstrated its value in web-facing businesses. With the digitalisation of the automotive industry, the focus in the industry is shifting towards software. For automotive embedded software to continuously improve, A/B experimentation is considered an important technique. However, the adoption of such a technique is not without challenge. In this paper, we present an architecture to enable A/B testing in automotive embedded software. The design addresses challenges that are unique to the automotive industry in a systematic fashion. Going from hypothesis to practice, our architecture was also applied in practice for running online experiments on a considerable scale. Furthermore, a case study approach was used to compare our proposal with state-of-practice in the automotive industry. We found our architecture design to be relevant and applicable in the efforts of adopting continuous A/B experiments in automotive embedded software.

The lack of labeled data is a major problem in both research and industrial settings since obtaining labels is often an expensive and time-consuming activity. In the past years, several machine learning algorithms were developed to assist and perform automated labeling in partially labeled datasets. While many of these algorithms are available in open-source packages, there is a lack of research that investigates how these algorithms compare to each other for different types of datasets and with different percentages of available labels. To address this problem, this paper empirically evaluates and compares seven algorithms for automated labeling in terms of their accuracy. We investigate how these algorithms perform in twelve different and well-known datasets with three different types of data, images, texts, and numerical values. We evaluate these algorithms under two different experimental conditions, with 10% and 50% labels of available labels in the dataset. Each algorithm, in each dataset for each experimental condition, is evaluated independently ten times with different random seeds. The results are analyzed and the algorithms are compared utilizing a Bayesian Bradley-Terry model. The results indicate that the active learning algorithms using the query strategies uncertainty sampling, QBC and random sampling are always the best algorithms. However, this comes with the expense of increased manual labeling effort. These results help machine learning practitioners in choosing optimal machine learning algorithms to label their data.

In practice, supervised learning algorithms require fully labeled datasets to achieve the high accuracy demanded by current modern applications. However, in industrial settings supervised learning algorithms can perform poorly because of few labeled instances. Semi-supervised learning (SSL) is an automatic labeling approach that utilizes complete labels to infer missing labels in partially complete datasets. The high number of available SSL algorithms and the lack of systematic comparison between them leaves practitioners without guidelines to select the appropriate one for their application. Moreover, each SSL algorithm is often validated and evaluated in a small number of common datasets. However, there is no research that examines what datasets are suitable for comparing different SSL algorihtms. The purpose of this paper is to empirically evaluate the suitability of the datasets commonly used to evaluate and compare different SSL algorithms. We performed a simulation study using twelve datasets of three different datatypes (numerical, text, image) on thirteen different SSL algorithms. The contributions of this paper are two-fold. First, we propose the use of Bayesian congeneric item response theory model to assess the suitability of commonly used datasets. Second, we compare the different SSL algorithms using these datasets. The results show that with except of three datasets, the others have very low discrimination factors and are easily solved by the current algorithms. Additionally, the SSL algorithms have overlapping 90% credible intervals, indicating uncertainty in the difference between the accuracy of these SSL models. The paper concludes suggesting that researchers and practitioners should better consider the choice of datasets used for comparing SSL algorithms.

A significant amount of research effort is put into studying machine learning (ML) and deep learning (DL) technologies. Real-world ML applications help companies to improve products and automate tasks such as classification, image recognition and automation. However, a traditional “fixed” approach where the system is frozen before deployment leads to a sub-optimal system performance. Systems autonomously experimenting with and improving their own behavior and performance could improve business outcomes but we need to know how this could actually work in practice. While there is some research on autonomously improving systems, the focus on the concepts and theoretical algorithms. However, less research is focused on empirical industry validation of the proposed theory. Empirical validations are usually done through simulations or by using synthetic or manually alteration of datasets. The contribution of this paper is twofold. First, we conduct a systematic literature review in which we focus on papers describing industrial deployments of autonomously improving systems and their real-world applications. Secondly, we identify open research questions and derive a model that classifies the level of autonomy based on our findings in the literature review. 

Randomised field experiments, such as A/B testing, have long been the gold standard for evaluating the value that new software brings to customers. However, running randomised field experiments is not always desired, possible or even ethical in the development of automotive embedded software. In the face of such restrictions, we propose the use of the Bayesian propensity score matching technique for causal inference of observational studies in the automotive domain. In this paper, we present a method based on the Bayesian propensity score matching framework, applied in the unique setting of automotive software engineering. This method is used to generate balanced control and treatment groups from an observational online evaluation and estimate causal treatment effects from the software changes, even with limited samples in the treatment group. We exemplify the method with a proof-of-concept in the automotive domain. In the example, we have a larger control (Nc = 1100) fleet of cars using the current software and a small treatment fleet (Nt = 38), in which we introduce a new software variant. We demonstrate a scenario that shipping of a new software to all users is restricted, as a result, a fully randomised experiment could not be conducted. Therefore, we utilised the Bayesian propensity score matching method with 14 observed covariates as inputs. The results show more balanced groups, suitable for estimating causal treatment effects from the collected observational data. We describe the method in detail and share our configuration. Furthermore, we discuss how can such a method be used for online evaluation of new software utilising small groups of samples.

With digitalization and with technologies such as software, data, and artificial intelligence, companies in the embedded systems domain are experiencing a rapid transformation of their conventional businesses. While the physical products and associated product sales provide the core revenue, these are increasingly being complemented with service offerings, new data-driven services, and digital products that allow for continuous value creation and delivery to customers. However, although there is significant research on digitalization and digital transformation, few studies highlight the specific needs of embedded systems companies and what it takes to transform from a traditional towards a digital company within business domains characterized by high complexity, hardware dependencies, and safety-critical system functionality. In this paper, we capture the difference between what constitutes a traditional and a digital company and we detail the typical evolution path embedded systems companies take when transitioning towards becoming digital companies.

In recent years, with the development of computation capability in devices, companies are eager to investigate and utilize suitable ML/DL methods to improve their service quality. However, with the traditional learning strategy, companies need to first build up a powerful data center to collect and analyze data from the edge and then perform centralized model training, which turns out to be inefficient. Federated Learning has been introduced to solve this challenge. Because of its characteristics such as model-only exchange and parallel training, the technique can not only preserve user data privacy but also accelerate model training speed. The method can easily handle real-time data generated from the edge without taking up a lot of valuable network transmission resources. In this paper, we introduce an approach to end-to-end on-device Machine Learning by utilizing Federated Learning. We validate our approach with an important industrial use case in the field of autonomous driving vehicles, the wheel steering angle prediction. Our results show that Federated Learning can significantly improve the quality of local edge models and also reach the same accuracy level as compared to the traditional centralized Machine Learning approach without its negative effects. Furthermore, Federated Learning can accelerate model training speed and reduce the communication overhead, which proves that this approach has great strength when deploying ML/DL components to various real-world embedded systems.

Artificial intelligence (AI) and machine learning (ML) are increasingly broadly adopted in industry. However, based on well over a dozen case studies, we have learned that deploying industry-strength, production quality ML models in systems proves to be challenging. Companies experience challenges related to data quality, design methods and processes, performance of models as well as deployment and compliance. We learned that a new, structured engineering approach is required to construct and evolve systems that contain ML/DL components. In this chapter, the authors provide a conceptualization of the typical evolution patterns that companies experience when employing ML as well as an overview of the key problems experienced by the companies that they have studied. The main contribution of the chapter is a research agenda for AI engineering that provides an overview of the key engineering challenges surrounding ML solutions and an overview of open items that need to be addressed by the research community at large.

With the increasing attention on Machine Learning applications, more and more companies are involved in implementing AI components into their software products in order to improve the service quality. With the rapid growth of distributed edge devices, Federated Learning has been introduced as a distributed learning technique, which enables model training in a large decentralized network without exchanging collected edge data. The method can not only preserve sensitive user data privacy but also save a large amount of data transmission bandwidth and the budget cost of computation equipment. In this paper, we provide a state-of-the-art overview of the empirical results reported in the existing literature regarding Federated Learning. According to the problems they expressed and solved, we then categorize those deployments into different application domains, identify their challenges and then propose six open research questions. 

Context: Nowadays, the hype around artificial intelligence is at its absolute peak. Large amounts of data are collected every second of the day and a variety of tools exists to enable easy analysis of data. In practice, however, making meaningful use of it is way more challenging. For instance, affected stakeholders often struggle to specify their information needs and to interpret the results of such analyses. Objective: In this study we investigate how to enable continuous monitoring of information needs, and the generation of knowledge and insights for various stakeholders involved in the lifecycle of software-intensive products. The overarching goal is to support their decision making by providing relevant insights related to their area of responsibility. Methods: We implement multiple monitoringand decision-making frameworks for six individual, real-world cases selected from three different platforms and covering four types of stakeholders. We compare the individual procedures to derive a generic process for instantiating such frameworks as well as a model to scale it up for multiple stakeholders. Results: For one, we discovered that information needs of stakeholders are often related to a limited subset of data sources and should be specified in stages. For another, stakeholders often benefit from sharing and reusing existing components among themselves in later phases. Specifically, we identify three types of reuse: (1) Data and knowledge, (2) tools and methods, and (3) concepts. As a result, key aspects of our model are iterative feedback and specification cycles as well as the reuse of appropriate components to speed up the instantiation process and maximize the efficiency of the model. Conclusion: Our results indicate that knowledge and insights can be generated much faster and stakeholders feel the benefits of the analysis very early on by iteratively specifying information needs and by systematically sharing and reusing knowledge, tools and concepts.

The impact of the Artificial Intelligence revolution is undoubtedly substantial in our society, life, firms, and employment. With data being a critical element, organizations are working towards obtaining high-quality data to train their AI models. Although data, data management, and data pipelines are part of industrial practice even before the introduction of ML models, the significance of data increased further with the advent of ML models, which force data pipeline developers to go beyond the traditional focus on data quality. The objective of this study is to analyze the impact of ML use cases on data pipelines. We assume that the data pipelines that serve ML models are given more importance compared to the conventional data pipelines. We report on a study that we conducted by observing software teams at three companies as they develop both conventional(Non-ML) data pipelines and data pipelines that serve ML-based applications. We study six data pipelines from three companies and categorize them based on their criticality and purpose. Further, we identify the determinants that can be used to compare the development and maintenance of these data pipelines. Finally, we map these factors in a two-dimensional space to illustrate their importance on a scale of low, moderate, and high.

Data pipelines play an important role throughout the data management process whether these are used for data analytics or machine learning. Data-driven organizations can make use of data pipelines for producing good quality data applications. Moreover, data pipelines ensure end-to-end velocity by automating the processes involved in extracting, transforming, combining, validating, and loading data for further analysis and visualization. However, the robustness of data pipelines is equally important since unhealthy data pipelines can add more noise to the input data. This paper identifies the essential elements for a robust data pipeline and analyses the trade-off between data pipeline robustness and complexity.

With the development and the increasing interests in ML/DL fields, companies are eager to apply Machine Learning/Deep Learning approaches to increase service quality and customer experience. Federated Learning was implemented as an effective model training method for distributing and accelerating time-consuming model training while protecting user data privacy. However, common Federated Learning approaches, on the other hand, use a synchronous protocol to conduct model aggregation, which is inflexible and unable to adapt to rapidly changing environments and heterogeneous hardware settings in real-world scenarios. In this paper, we present an approach to real-time end-to-end Federated Learning combined with a novel asynchronous model aggregation protocol. Our method is validated in an industrial use case in the automotive domain, focusing on steering wheel angle prediction for autonomous driving. Our findings show that asynchronous Federated Learning can significantly improve the prediction performance of local edge models while maintaining the same level of accuracy as centralized machine learning. Furthermore, by using a sliding training window, the approach can minimize communication overhead, accelerate model training speed and consume real-time streaming data, proving high efficiency when deploying ML/DL components to heterogeneous real-world embedded systems.

A/B testing is gaining attention in the automotive sector as a promising tool to measure casual effects from software changes. Different from the web-facing businesses, where A/B testing has been well-established, the automotive domain often suffers from limited eligible users to participate in online experiments. To address this shortcoming, we present a method for designing balanced control and treatment groups so that sound conclusions can be drawn from experiments with considerably small sample sizes. While the Balance Match Weighted method has been used in other domains such as medicine, this is the first paper to apply and evaluate it in the context of software development. Furthermore, we describe the Balance Match Weighted method in detail and we conduct a case study together with an automotive manufacturer to apply the group design method in a fleet of vehicles. Finally, we present our case study in the automotive software engineering domain, as well as a discussion on the benefits and limitations of the A/B group design method.

Frequentist statistical methods, such as hypothesis testing, are standard practices in studies that provide benchmark comparisons. Unfortunately, these methods have often been misused, e.g., without testing for their statistical test assumptions or without controlling for familywise errors in multiple group comparisons, among several other problems. Bayesian data analysis (BDA) addresses many of the previously mentioned shortcomings but its use is not widely spread in the analysis of empirical data in the evolutionary computing community. This article provides three main contributions. First, we motivate the need for utilizing BDA and provide an overview of this topic. Second, we discuss the practical aspects of BDA to ensure that our models are valid and the results are transparent. Finally, we provide five statistical models that can be used to answer multiple research questions. The online Appendix provides a step-by-step guide on how to perform the analysis of the models discussed in this article, including the code for the statistical models, the data transformations, and the discussed tables and figures. 

Data collected from in-service products play an important role in enabling software-intensive embedded systems suppliers to embrace data-driven practices. Data can be used in many different ways such as to continuously learn and improve the product, enhance post-deployment services, reduce operational cost or create a better user experience. While there is no shortage of possible use cases leveraging data from in-service products, software-intensive embedded systems companies struggle to continuously collect data from their in-service products. Often, data collection is done in an ad-hoc way and targeting specific use cases or needs. Besides, few studies have investigated data collection challenges in relation to the data dimensions, which are the minimum set of quantifiable data aspects that can define software-intensive embedded product data from a collection point of view. To help address data collection challenges, and to provide companies with guidance on how to improve this process, we conducted a case study at a large multinational telecommunications supplier focusing on data characteristics and collection challenges from the Radio Access Networks (RAN) products. We further investigated the relations of these challenges to the data dimensions to increase our understanding of how data dominions contribute to the challenges.

Federated Learning, as a distributed learning technique, has emerged with the improvement of the performance of IoT and edge devices. The emergence of this learning method alters the situation in which data must be centrally uploaded to the cloud for processing and maximizes the utilization of edge devices' computing and storage capabilities. The learning approach eliminates the need to upload large amounts of local data and reduces data transfer latency with local data processing. Since the Federated Learning technique does not require centralized data for model training, it is better suited to edge learning scenarios in which nodes have limited data. However, despite the fact that Federated Learning has significant benefits, we discovered that companies struggle with integrating Federated Learning components into their systems. In this paper, we present case study research that describes reasons why companies anticipate Federated Learning as an applicable technique. Secondly, we summarize the services that a complete Federated Learning system needs to support in industrial scenarios and then identify the key challenges for industries to adopt and transition to Federated Learning. Finally, based on our empirical findings, we suggest five criteria for companies implementing reliable Federated Learning systems.

The adoption of continuous software engineering practices such as DevOps (Development and Operations) in business operations has contributed to significantly shorter software development and deployment cycles. Recently, the term MLOps (Machine Learning Operations) has gained increasing interest as a practice that brings together data scientists and operations teams. However, the adoption of MLOps in practice is still in its infancy and there are few common guidelines on how to effectively integrate it into existing software development practices. In this paper, we conduct a systematic literature review and a grey literature review to derive a framework that identifies the activities involved in the adoption of MLOps and the stages in which companies evolve as they become more mature and advanced. We validate this framework in three case companies and show how they have managed to adopt and integrate MLOps in their large-scale software development companies. The contribution of this paper is threefold. First, we review contemporary literature to provide an overview of the state-of-the-art in MLOps. Based on this review, we derive an MLOps framework that details the activities involved in the continuous development of machine learning models. Second, we present a maturity model in which we outline the different stages that companies go through in evolving their MLOps practices. Third, we validate our framework in three embedded systems case companies and map the companies to the stages in the maturity model. 

Machine learning and deep learning techniques are becoming increasingly popular and critical for companies as part of their systems. However, although the development and prototyping of ML/DL systems are common across companies, the transition from prototype to production-quality deployment models are challenging. One of the key challenges is how to determine the selection of an optimal architecture for AI deployment. Based on our previous research, and to offer support and guidance to practitioners, we developed a framework in which we present five architectural alternatives for AI deployment ranging from centralized to fully decentralized edge architectures. As part of our research, we validated the framework in software-intensive embedded system companies and identified key challenges they face when deploying ML/DL models. In this paper, and to further advance our research on this topic, we identify factors that help practitioners determine what architecture to select for the ML/D L model deployment. For this, we conducted a follow-up study involving interviews and workshops in seven case companies in the embedded systems domain. Based on our findings, we identify three key factors and develop a framework in which we outline how prioritization and trade-offs between these result in certain architecture. The contribution of the paper is threefold. First, we identify key factors critical for AI system deployment. Second, we present the architecture selection framework that explains how prioritization and trade-offs between key factors result in the selection of a certain architecture. Third, we discuss additional factors that may or may not influence the selection of an optimal architecture.

Since the advent of mobile computing and IoT, a large amount of data is distributed around the world. Companies are increasingly experimenting with innovative ways of implementing edge/cloud (re)training of AI systems to exploit large quantities of data to optimize their business value. Despite the obvious benefits, companies face challenges as the decision on how to implement edge/cloud (re)training depends on factors such as the task intent, the amount of data needed for (re)training, edge-to-cloud data transfer, the available computing and memory resources. Based on action research in a software-intensive embedded systems company where we study multiple use cases as well as insights from our previous collaborations with industry, we develop a generic framework consisting of five architectural alternatives to deploy AI on the edge utilizing transfer learning. We validate the framework in four additional case companies and present the challenges they face in selecting the optimal architecture. The contribution of the paper is threefold. First, we develop a generic framework consisting of five architectural alternatives ranging from a centralized architecture where cloud (re)training is given priority to a decentralized architecture where edge (re)training is instead given priority. Second, we validate the framework in a qualitative interview study with four additional case companies. As an outcome of validation study, we present two variants to the architectural alternatives identified as part of the framework. Finally, we identify the key challenges that experts face in selecting an ideal architectural alternative.

Nowadays, machine learning (ML) is an integral component in a wide range of areas, including software analytics (SA) and business intelligence (BI). As a result, the interest in custom ML-based software analytics and business intelligence solutions is rising. In practice, however, such solutions often get stuck in a prototypical stage because setting up an infrastructure for deployment and maintenance is considered complex and time-consuming. For this reason, we aim at structuring the entire process and making it more transparent by deriving an end-to-end framework from existing literature for building and deploying ML-based software analytics and business intelligence solutions. The framework is structured in three iterative cycles representing different stages in a model’s lifecycle: prototyping, deployment, update. As a result, the framework specifically supports the transitions between these stages while also covering all important activities from data collection to retraining deployed ML models. To validate the applicability of the framework in practice, we compare it to and apply it in a real-world ML-based SA/BI solution.

Companies across domains are rapidly engaged in shifting computational power and intelligence from centralized cloud to fully decentralized edges to maximize value delivery, strengthen security and reduce latency. However, most companies have only recently started pursuing this opportunity and are therefore at the early stage of the cloud-to-edge transition. To provide an overview of AI deployment in the context of edge/cloud/hybrid architectures, we conduct a systematic literature review and a grey literature review. To advance understanding of how to integrate, deploy, operationalize and evolve AI models, we derive a framework from existing literature to accelerate the end-to-end deployment process. The framework is organized into five phases: Design, Integration, Deployment, Operation and Evolution. We make an attempt to analyze the extracted results by comparing and contrasting them to derive insights. The contribution of the paper is threefold. First, we conduct a systematic literature review in which we review the contemporary scientific literature and provide a detailed overview of the state-of-the-art of AI deployment. Second, we review the grey literature and present the state-of-practice and experience of practitioners while deploying AI models. Third, we present a framework derived from existing literature for the end-to-end deployment process and attempt to compare and contrast SLR and GLR results.

Over the past 15 years, A/B testing has been a critical tool for accurate prioritization of development efforts in online and web-facing companies. As automotive companies progress on their digitalization process, A/B testing and other experimentation techniques start to be adopted. However, specific characteristics of the automotive software industry create additional challenges to the successful adoption of A/B testing. Recently, research has been conducted to investigate the challenges and opportunities for experimentation techniques in the automotive and more generally in the embedded systems domain. However, despite the collaboration with industry, previous research was based on either hypothesized or toy scenarios in companies seeking, but not yet running experimentation. Utilizing a case study method, we investigate the challenges of adopting A/B testing in two large-scale automotive companies that are currently running or preparing for their first A/B testing. The contribution of this paper is two-fold. First, we present our main findings in terms of the challenges of real A/B testing iterations in automotive vehicles. Second, we present the current, potential solutions and lessons learned from applying A/B testing in the automotive domain.

In recent years, the application of artificial intelligence (AI) has become an integral part of a wide range of areas, including software engineering. By analyzing various data sources generated in software engineering, it can provide valuable insights into customer behavior, product performance, bugs and errors, and many more. In practice, however, AI for software analytics and business intelligence often gets stuck in a prototypical stage and the results are rarely used to make decisions based on data. To understand the underlying root causes of this phenomenon, we conduct both an explanatory case study and a survey on the challenges of realizing and utilizing artificial intelligence in the context of software-intensive businesses. As a result, we identify a vicious circle that prevents practitioners from moving from prototypical analytics to continuous and productively usable software analytics and business intelligence based on AI.

Labeling is a cornerstone of supervised machine learning. However, in industrial applications, data is often not labeled, which complicates using this data for machine learning. Although there are well-established labeling techniques such as crowdsourcing, active learning, and semi-supervised learning, these still do not provide accurate and reliable labels for every machine learning use case in the industry. In this context, the industry still relies heavily on manually annotating and labeling their data. This study investigates the challenges that companies experience when annotating and labeling their data. We performed a case study using a semi-structured interview with data scientists at two companies to explore their problems when labeling and annotating their data. This paper provides two contributions. We identify industry challenges in the labeling process, and then we propose mitigation strategies for these challenges.

Data pipelines involve a complex chain of interconnected activities that starts with a data source and ends in a data sink. Data pipelines are important for data-driven organizations since a data pipeline can process data in multiple formats from distributed data sources with minimal human intervention, accelerate data life cycle activities, and enhance productivity in data-driven enterprises. However, there are challenges and opportunities in implementing data pipelines but practical industry experiences are seldom reported. The findings of this study are derived by conducting a qualitative multiple-case study and interviews with the representatives of three companies. The challenges include data quality issues, infrastructure maintenance problems, and organizational barriers. On the other hand, data pipelines are implemented to enable traceability, fault-tolerance, and reduce human errors through maximizing automation thereby producing high-quality data. Based on multiple-case study research with five use cases from three case companies, this paper identifies the key challenges and benefits associated with the implementation and use of data pipelines.

Artificial Intelligence is becoming increasingly popular with organizations due to the success of Machine Learning and Deep Learning techniques. Using these techniques, data scientists learn from vast amounts of data to enhance behaviour in software-intensive systems. Despite the attractiveness of these techniques, however, there is a lack of systematic and structured design process for developing ML/DL models. The study uses a multiple-case study approach to explore the different activities and challenges data scientists face when developing ML/DL models in software-intensive embedded systems. In addition, we have identified seven different phases in the proposed design process leading to effective model development based on the case study. Iterations identified between phases and events which trigger these iterations optimize the design process for ML/DL models. Lessons learned from this study allow data scientists and engineers to develop high-performance ML/DL models and also bridge the gap between high demand and low supply of data scientists.

Continuous experimentation (CE) refers to a group of practices used by software companies to rapidly assess the usage, value and performance of deployed software using data collected from customers and the deployed system. Despite its increasing popularity in the development of web-facing applications, CE has not been discussed in the development process of business-to-business (B2B) mission-critical systems.

We investigated in a case study the use of CE practices within several products, teams and areas inside Ericsson. By observing the CE practices of different teams, we were able to identify the key activities in four main areas and inductively derive an experimentation process, the HURRIER process, that addresses the deployment of experiments with customers in the B2B and with mission-critical systems. We illustrate this process with a case study in the development of a large mission-critical functionality in the Long Term Evolution (4G) product. In this case study, the HURRIER process is not only used to validate the value delivered by the solution but to increase the quality and the confidence from both the customers and the R&D organization in the deployed solution. Additionally, we discuss the challenges, opportunities and lessons learned from applying CE and the HURRIER process in B2B mission-critical systems.

Machine Learning (ML) and Artificial Intelligence (AI) have increasingly gained attention in research and industry. Federated Learning, as an approach to distributed learning, shows its potential with the increasing number of devices on the edge and the development of computing power. However, most of the current Federated Learning systems apply a single-server centralized architecture, which may cause several critical problems, such as the single-point of failure as well as scaling and performance problems. In this paper, we propose and compare four architecture alternatives for a Federated Learning system, i.e. centralized, hierarchical, regional and decentralized architectures. We conduct the study by using two well-known data sets and measuring several system performance metrics for all four alternatives. Our results suggest scenarios and use cases which are suitable for each alternative. In addition, we investigate the trade-off between communication latency, model evolution time and the model classification performance, which is crucial to applying the results into real-world industrial systems.

The collection of high-quality data provides a key competitive advantage to companies in their decision-making process. It helps to understand customer behavior and enables the usage and deployment of new technologies based on machine learning. However, the process from collecting the data, to clean and process it to be used by data scientists and applications is often manual, non-optimized and error-prone. This increases the time that the data takes to deliver value for the business. To reduce this time companies are looking into automation and validation of the data processes. Data processes are the operational side of data analytic workflow.

DataOps, a recently coined term by data scientists, data analysts and data engineers refer to a general process aimed to shorten the end-to-end data analytic life-cycle time by introducing automation in the data collection, validation, and verification process. Despite its increasing popularity among practitioners, research on this topic has been limited and does not provide a clear definition for the term or how a data analytic process evolves from ad-hoc data collection to fully automated data analytics as envisioned by DataOps.

This research provides three main contributions. First, utilizing multi-vocal literature we provide a definition and a scope for the general process referred to as DataOps. Second, based on a case study with a large mobile telecommunication organization, we analyze how multiple data analytic teams evolve their infrastructure and processes towards DataOps. Also, we provide a stairway showing the different stages of the evolution process. With this evolution model, companies can identify the stage which they belong to and also, can try to move to the next stage by overcoming the challenges they encounter in the current stage.