Self-Adaptive Systems: A Systematic Literature Review Across Categories and Domains

We used the following research questions to drive our data collection and analysis:

RQ1

What is the current state of the art in self-adaptive systems?

RQ2

How has the state of the art evolved over time?

RQ3

Which are the application domains of self-adaptive systems over time?

To select literature to include in our systematic review on self-adaptive systems, we used the dblp computer science bibliography (dblp) database as a starting point, effectively eliminating non-computer science literature from the process. Dblp contains historical snapshots of the database enabling reproducibility of results. All results in this review were produced using the snapshot file dblp-2020-03-02.xml.¹¹1https://dblp.org/xml/release/dblp-2020-03-02.xml.gz

We chose dblp as the starting point of our analysis since it provides a consistent format for all articles indexed in the database which allowed us to employ consistent search criteria (such as lower-case vs. upper-case) across content published by different publishers. Dblp’s focus on “major computer science publications”²²2https://dblp.org/faq/What+is+dblp.html gives us a narrower focus than a more general search engine and a finite number of search results that does not vary on a daily basis, as it would for example on Google Scholar. We acknowledge that using a different database as a starting point would have resulted in a different set of papers, see Section 5 for a discussion of the corresponding trade-offs.

The methodology is broken down into five main stages which combine automated and manual processes to keep the required amount of manual work (i.e., reading abstracts and papers) manageable while ensuring the quality of the selection process: Pre-filtering, word frequency filtering, venue selection, abstract reading, and snowballing to reduce a first subset of 35,903 publications to 293 publications. Figure 1 shows a high-level overview of the process.

According to our literature review, the earliest reference to self-adaptive systems was in 1990. The period of 1990-2003 was the phase before ‘The Vision of Autonomic Computing’ [2], was published. This was where the field was in its beginning stages.

References in this phase contained theoretical and model based papers such as a model for dynamic change management [23], self-stabilizing real-time rule based systems [24], and convergence for self-stabilizing systems [25]. These papers were focused on presenting a theoretical model or proving a proposed theorem to advance self-adaptive systems theory.

By 1998, practical implementations were seen such as architecture based run-time software evolution [1], self-supervised category detection [26], and self-adaptive control software [27]. The papers were among the first to discuss terms like evolution, self-supervision, control theory, and run-time design in the context of self-adaptive systems. Thirty years later, these terms are commonplace and part of the general understanding of the field.

Tables 9 and 10 characterize this time period in terms of the prevalent paper categories and application domains.

In the subsequent years after 2002, two seminal works were produced that formed the foundations of self-adaptive systems. In 2003, The Vision of Autonomic Computing [2] was published which kick-started the field of Autonomic Computing. The paper presented a grand challenge of self-configuration, self-optimization, self-healing, and self-protection in computing systems. It foretasted the rising need for these systems in the modern age. Shortly after this, an implementation of these concepts was developed, known as the RAINBOW framework [28]. This framework was architecture based and extensible, meaning existing architectures could be imbued with self-adaptive properties. To this day, the RAINBOW framework is used as a standard, test bed, extensible tool in self-adaptive systems research.

Many papers were published during this time period that were in the perspective category (42%). As the excitement of a new field grew, many researchers sought to publish their thoughts and ideas as to how the field should and could develop. During this period, there were fewer practical demonstrations and implementations compared to future time periods as the field had to have time to grow and mature.

This application domain distribution of this period was skewed towards review papers. Of the eight review papers, four were on autonomic computing [2, 13, 11, 12] and four were on self-* properties [29, 30, 31, 32]. The key messages of these papers were that a new challenge was emerging in the field of autonomic computing and self-* computing. Due to the rise in system complexity, it was necessary to develop a new kind of system that was self-adaptive. These papers also envisioned what such realized systems might look like.

The most immediate application domain for self-adaptive systems was web services. At the time, technologies such as IaaS and IoT had not fully emerged yet. Web services were a prominent tool used in software engineering.

The improvement of the system administrator role was a focus area of research by improving collaboration and coordination, rehearsal and planning, maintaining situation awareness, and managing multitasking, interruptions, and diversions [33]. Utility functions emerged as a potential metric to solve self-adaptive problems [34]. In later years, this would prove to be true. The automatic management of web services was an important test bed to develop self-adaptive algorithms and theory [35, 36, 37].

There was a major parallel between autonomic computing, self-* computing, and biological systems, and a branch of research was formed to gain inspiration from nature to bring to these systems [38, 39, 40]. As a precursor to the cloud based systems of the present day, load balancing was an area of research focus. Work forecasting [41], scheduling algorithms [42], and managing system level properties [43] were all part of the groundwork of this research area, as was the placement of replicants in an edge computing scenario [44].

IoT and IaaS papers were present during this time. Concepts like autonomous deployment [45], generic control frameworks [46], and self-healing distributed architectures [47] were explored, as was the concept of self-aware systems [48] in general. These two fields would emerge as strong research fields in later years.

Other research areas include Robotics [49], Automotive [50], Computer [51], Software Stack [52], Networking [53, 54], Software Engineering [55], and Intelligence Surveillance Reconnaissance [56].

Tables 11 and 12 characterize this time period in terms of the prevalent paper categories and application domains, confirming the large number of perspective papers.

The distribution of the years 2006-2010 showed a more even spread across the categories. Technological, Perspective, and Methodology papers were evenly spread. There was still a high demand for perspective type papers but enough time had passed since 2003 for technology and methodology papers to emerge.

Web services was still the highest polling application domain but in this period IoT papers surfaced as the second most frequent domain. Web services in this phase focused on the principles and ideas of autonomic and self-adaptive systems developed in the previous phase. These manifested initially in new languages [57], utility functions [57], and distributed solutions as opposed to centralized solutions [58]. There was a clear focus on run time instead of design time solutions [59, 60] as well as making systems dynamic in their configuration and adaptation [61, 62, 63]. Self-healing was an important part of web services research during this time, highlighting the use in workarounds [64] and composition cycles [65]. The RAINBOW framework was evaluated on a signature case study: Znn.com to assess the effectiveness of self-adaptation with good results [66]. In the latter half of this period, the focus of web services grew to become more holistic with concepts such as architectural self-reconfiguration and self-tuning [67, 68], behavioral adaptations [69], expansion into more self-* properties [70], and adaption logic based approaches [71]. The emphasis was on adaptation of the entire system as opposed to individual components.

IoT papers were initially focused on solvers [72, 73] with a key driver being dynamic behavior [74, 75]. Fault tolerance [74] and multi-agent models [72] were important themes to establish in IoT due to their distributed and always-on nature. A tool called RELAX was developed to handle requirements and uncertainty in a smart home self-adaptive system domain [76]. Requirements and uncertainty are important in any self-adaptive system, not just IoT systems [77]. The requirements of the system determine the goals of the adaptations. Appropriately managing uncertainty gives a system flexibility and dynamism. As with web services, a holistic approach to self-adaptation using architectural self-reconfiguration was a part of the research [78]. Adaption logic in IoT was addressed using heuristics in availability and response time [79] and in a tool called FUSION [80]. The research efforts of IoT compared with web services in this phase share a parallel. Initially, solvers and languages were created and tested, then various research concepts were explored like distributed solutions for web services or fault tolerance for IoT. Then, a shift in focus to more holistic approaches to self-adaptability were emphasized such as architectural self-reconfiguration, or the adaptation logic.

This phase was the first time robotics was seen as an application domain with seven papers. Robotics represented a high potential area for the application of self-adaptive systems. By their nature, robots are intended to automatically do the tasks of humans and so self-adaptive robots are a natural progression of the goals. Requirements and modelling [81, 82] were the early emphasis with a progression to adaptable software architectures [82, 83] and self-organized and distributed systems [84]. A reference model to compare adaptation approaches was developed [85]. This started to become necessary as the popularity of the field gave rise to multiple different adaptation techniques. Learning and planning was a component of the research of this phase [86, 87]. Using reinforcement learning, planning and architecture self-management was explored [86]. The learning research theme was in its early stages in this phase. In future years, learning would prove to be key to widely used tools like machine and deep learning.

This phase also gave rise to nine review papers. Compared to the previous phase, review papers made up a smaller proportion of the total papers in the group. The reviews were conducted on self-organization [88, 89], autonomic computing [90, 14, 91], self-healing [18], and self-adaptation [15, 92], as well as a general overview of self-* properties [93]. The common theme of this work was to highlight the current work and future needs of the field. One remark confirms the trend highlighted above of needing a more holistic approach to self-adaptation: ”but what is missing is an holistic approach focusing explicitly on providing autonomic properties” [91]. Written in 2009, it perhaps explains the trend toward holistic solutions to self-adaptive systems in the literature.

This phase showed a variety of other application domains. Decentralized solutions were popular in automated traffic control [94], e-Commerce [95], water networks [96], service oriented systems [97], and load balancing solutions [98]. Learning was a prominent theme for automotive [99], traffic management control [100], and e-Commerce [101]. The ability for a system to adapt its adaptation logic was a strong area of focus for mobile systems[102], ISR [103], e-Commerce [104], software engineering applications [105, 106], and systems on chip [107]. Evaluation [108, 109], self-organization [110, 111, 112], self-protection [113], self-healing [114], resource allocation [115], reflection [116], as well as dynamic solutions [117] were all explored during this time. Consistent with the theme of holistic solutions, generic architectures were developed in speech recognition tools [118] as well as generic frameworks such as SASSY [119]. Other application domains not listed include UML [120], Holonic Systems [121], Video Encoding [122], task scheduling [123], and Physics [124].

Tables 13 and 14 characterize this time period in terms of the prevalent paper categories and application domains, underlining that papers in the ‘Methodology’ category and about Web Services were becoming increasingly dominant.

The first half of the 2010s showed a bimodal result, with technological and methodology papers being the most frequent. The proportion of perspective papers decreased, perhaps attributed to the saturation of them in previous time periods.

Web services still dominated the domain distribution, followed by robotics and networking. IoT still had a high ranking with seven papers. This was the first time period where IaaS made it in to the top domains. This is in line with the timeline of cloud based services as they rose to prominence.

Web services papers in this time period displayed a slightly different flavor to the previous time periods. Forms of validation were more popular in this phase. Performance [125] and integration [126] testing, probabilistic model checking [127, 128, 129], quality assurance [130], and evaluation [131] indicate that more emphasis was now being put on verifying the outcomes of self-adaptive solutions. It was not enough to claim that a system was self-adaptive, but the claims had to be backed up and tested. Frameworks formed part of the contributions to the research. With the push towards more holistic self-adaptive solutions, this was a natural progression. Multi-model [132], dynamic allocation [133], monitoring [134], and behavioral [135] frameworks highlight the variety of holistic generic approaches being explored. The RAINBOW framework still held influence during this period, with a framework called REFRACT extending RAINBOW to target fault avoidance [136]. Planning was a key component of this phase – automatic reconfiguration plans [137], adapting manager optimization [138], and plan generation techniques [139] all highlight the continued importance of forecasting in self-adaptive systems. Distributed techniques [140], fault localization [141], and architecture based self-adaptation [142] all were continued self-adaptive themes from the previous phase.

Robotics papers also shared an emphasis on building generic frameworks for self-adaptation. Architectural compilers [143], reference models [144], and testing frameworks [145] highlight some of the work done here. There was a continued focus on dynamic [146] and run-time [147] applications as well as verification of systems [148]. The modelling of uncertainty became a strong focus during this phase. As decisions in the system are pushed from design time to run time, the amount of possible outcomes for the system dramatically increases. The behavior of the system also becomes non-deterministic. Verification [149], consequence modelling [150], and latency modelling [128] all attest to this effort. A strong research theme of the robotics domain in self-adaptive systems is goal modelling. The goals of a robot and associated utility functions are highly important to successful behavior. Dealing with fuzzy goals [151], interactions [150], and learning [152] all contributed to this focus.

Networking was a domain that was well represented during this time period. With the explosion of the Internet and always-on devices, networking was a ripe domain to apply self-adaptive principles to. Consistent with the theme of frameworks during this time period, mathematical [153], scheduling [154], sensor modelling [155], and testing [156] frameworks were created. There was a similar emphasis on validation [157] and fault tolerance [158]. The RAINBOW framework was again used as an exemplar during this phase. It was applied to manage and monitor highly populated networks of devices [159]. Consistent with the themes from previous phases, self-organization [160] and self-reconfiguration [161, 162] were popular research areas in networking.

The IoT domain shares a high overlap with the networking domain as IoT solutions are essentially localized networks. The trends of the IoT research in this phase is consistent with the themes of generic frameworks [163, 164, 165, 166, 167] and validation [168, 169].

The IaaS domain entered the top domains in this phase. In the second half of the decade its popularity would explode. During this time, IaaS papers focused on regression testing [170], control theory [171], decision making [172, 173], transaction management [174], and on benchmarking and elasticity [175]. These themes would later on be developed and expanded.

There were only six reviews in this time period, the lowest proportion of any time period so far. This could be due to the reduced need because of the ongoing work. Reviews in this period focused on self-healing [19, 20], self-adaptation [7, 16], self-protection [21], and on control engineering approaches to self-adaptive system design [176].

Frameworks in ISR [177, 178], service oriented systems [179, 180, 181], software engineering [182, 183], and mobile systems [184, 185] highlight the trend of the theme in developing holistic solutions to self-adaptive systems. Continued trends are requirements [186, 187], dynamic solutions [188], multi-agent systems [189, 190], and utility [191] as well as reliability [192], with more abstract ideas like systems evaluation [193], uncertainty handling [194, 195], and feedback loops [196, 197] now getting covered.

Bio-inspired approaches were again present in this phase. These approaches have the understanding that self-adaptive systems are much like biological systems and that there is much inspiration to draw from nature. Papers discuss chemically inspired architectures for reusable models [198], as well as cloud based applications inspired by biological principles [199] and multi-objective control for self-adaptive software design [200]. These biological inspired approaches are a potential growth area for self-adaptive systems.

Other domain applications not listed include e-Commerce [201], UML [202], automotive [203], water networks [204], automated traffic management [205], fault recovery [206], video encoding [207], application containers [208], and human participation [209].

Tables 15 and 16 characterize this time period in terms of the prevalent paper categories and application domains, showing a large number of papers in the categories ‘Technological’ and ‘Methodology’, again dominated by the application area of Web Services.

The second half of the 2010s displayed a similar distribution to the first half, with technological and methodology papers being most frequent, followed by analytical, perspective, and empirical. This is not surprising as the developments of the field were at a comparable maturation stage.

The distribution of the domains during this period shows an interesting trend. Web services are no longer the most frequent domain, rather IoT and IaaS are the most frequent domains. There is a distinct trend of these cloud based technologies from obscurity (1990-2003) to niche (2003-2010) to growth (2011-2015) to now in the 2020s where cloud based services are mainstream. This trend is reflected in the rise of IoT and IaaS in the domain distribution of the papers.

IoT has become a popular application domain in the last five years of the 2010s. The number of devices with access to the Internet has increased exponentially in quantity but also variety. Devices are not just limited to phones and computers but extend to watches, cars, buildings, sensors, and more. The research in this phase gave rise to a number of exemplars in IoT. Exemplars can be generic such as artifacts or address specific self-adaptive problems. They are used as a demonstration of a working solution in the problem space. This was the period of time where exemplars began to be widely seen and used. DeltaIoT [210], an evaluation exemplar and DingNet [211], a simulation exemplar highlight the work done. Carrying on from the work done in the previous time periods, frameworks and generic architectures were seen in HAFLoop [212], decentralized approaches [213, 214], and modelling frameworks [215]. Behavioral modelling of IoT behavior was a continued trend [216, 217, 218, 219]. Specifically, emergent behaviors were explored [220]. Emergent behavior is an important concept in self-adaptive systems. These behaviors are the byproduct of allowing decisions to be made at run time. When this occurs, the system may display new behavior not previously conceived or seen before. The appropriate handling of these behaviors is important to a large scale self-adaptive system like IoT. Common research focuses like new languages [221], evaluation and testing [222, 223], modelling [224], learning [225, 226, 227], recovery [228], uncertainty [229], and integration [230] were seen in this phase. A general review of self-improving system integration can be found in [231] and industrial experience reports in [232, 233].

IaaS became more of a focus during this period. As it went further into the decade, cloud based solutions became more and more common in business and hence in research. It became cheaper to rent out infrastructure in the cloud and outsource maintenance costs than to handle everything in-house. This is also reflected in research focusing on concurrent approaches [234, 235] or hierarchical systems [236].

From the previous time period, the research areas of testing [237], control theory [213], decision making [238, 239, 240], and elasticity [225] in IaaS were expanded on. New areas of research, such as trust [241], structural and parametric adaptation [242], monitoring [224, 243], modelling [244, 245], and service level maintenance [246] were established in the field. Trust in self-adaptive systems is an important concept. Even if the self-adaptation loops are robust and effective, without establishing trust for the system, using these systems in large scale or critical environments is infeasible. This has to to with the understood error rate of the self-adaptive system and the tolerance of the user. In some cases it may be acceptable to have a 20% error rate in a non-critical scenario but for another critical scenario like a Defence setting, even a 5% error rate may not be acceptable given the possible consequences. The research efforts in self-adaptive systems are usually split between structural and parametric adaptation. Structural adaptation involves modifying or improving the components of the system whereas parametric adaptation involves optimizing the configurable parameters of the system, leaving the components unchanged. Addressing both of the styles at once is an area of need and potential [242].

The research into web services has benefited from a 20 plus year build up. Consistent with the trend of this time period, exemplars were used to demonstrate the capabilities of self-adaptive web services using TCP communication [247]. Multi-agent systems [248], uncertainty [249], planning [250], models [251, 252], and programming concepts [253, 254] were all continued research themes into web services. The state of web services after 20 plus years has moved from foundational theory to generic frameworks and to exemplars. Even despite this trend, the various research themes are still being explored and mined for use after 20 years which indicates that there is more to learn.

Cyber-Physical Systems (CPS) are a combination of computation, networking, and physical processes where a physical component is controlled by a chip or software component. This time period is the first time CPS are seen. This indicates that they are a relatively new research area to self-adaptive systems. The dominating application (24%) to CPS is energy and the dominant adaptation mechanism is MAPE-K [255]. A new language, Adaptive CSP was developed to support compositional verification of systems [256]. Continuing with the trend of exemplars in this phase, DARTSim represents a simulation of UAVs on a reconnaissance mission communicating via TCP [257]. According to [258], a central concept in these systems is homeostasis, the capacity to maintain an operational state despite run-time uncertainty. This is addressed by four principles: collaborative sensing, faulty component isolation from adaptation, enhancing mode switching, and adjusting guards in mode switching. CPS are naturally employed in safety-critical environments as their small nature allows them to be embedded into any physical tool. The successful integration [259] and the tracability of these components [260] are critical to the field.

Continuing on with the last time period, bio inspired approaches were seen in this time period, addressing emergent behavior [261] and artificial DNA [262]. Security was a focus of this theme with guarantees [263] and verifications [264] being explored. The self-protection aspect of self-adaptive systems has been sprinkled amongst the time periods (with works focusing on trust [265, 266, 267] and on situational awareness [268]), however as these systems gain traction and popularity, there will be an increased need to secure these systems in the same fashion as micro-transactions are secured in financial institutions. It would not be surprising if in the next decade, security became a more prominent application domain.

There was a paper in this time period on smart factory or industry 4.0 [269]. Seen as the next progression in industrial activity, this application domain has potential to grow going in to the next decade. Exemplars were again seen in this period across other domains using architectural self-healing and self-optimization [270].

The frequencies of robotics and networking decreased in this phase. This could be because they are less popular or that there is some overlap between these domains and the top two domains, IoT and IaaS. The second reason is more likely. Planning [271, 272, 273, 274], testing [275], fault tolerance [276], and uncertainty [277, 278] all highlight common research trends seen before in the timelines, as is model-predictive control [279]. Mobile systems most likely also share the same similarities with overlap as networking and robotics to IoT and IaaS. They have been a consistent theme across the timelines and have a presence in this one with dynamic decisions [280], input space mapping [281] and emotion measurement [282]. A review of self-adaptive systems in the context of mobile systems is given in [283], one on monitoring self-adaptive applications within edge computing frameworks in [284], and one on learning in self-adaptive systems in [285].

The automotive application domain increased in this time period compared to the previous time period. This may be explained by the new found viability of smart cars and self-driving cars in recent years. Key trends for this domain were adaptive, scalable, and robust systems. These systems are proactively aware of latency and can act in swarms [286, 287, 288, 289, 290].

ISR has had a consistent presence across the timelines. Resilience [291], goal theory [292], control theory [293], and assurance [294] highlight the research efforts in this time period. ISR is an important application domain to self-adaptive systems. It enables real time situational awareness and allows analysts to make decisions based off current and useful information. In a Defence context, generating this intelligence from data is extremely important to the decision makers.

Other application domains included clonal plasticity [295], smart travels [296], agriculture [297], UML [298], system on chip [299], MAPE-K [300], traffic management [301], holonic systems [302], and managing support of recoonfigurable software components [303].

Tables 17 and 18 characterize this time period in terms of the prevalent paper categories and application domains, underlining the focus on IoT and IaaS in recent years.

In this section, we briefly revisit the research questions set out at the beginning of this review to summarize our findings.

RQ1 What is the current state of the art in self-adaptive systems? The current state of the art in self-adaptive systems is focused on developing methodologies and technology in the area of cloud-based services, such as IoT and IaaS. Although research on self-adaptive systems tends to be diverse, empirical and analytical research is currently playing a smaller role, as are other application domains. The importance of self-adaptive systems is rapidly growing in areas such as bio-inspired approaches, security, and cyber physical systems.

RQ2 How has the state of the art evolved over time? In the 1990s, research on self-adaptive systems started with theoretical and model based papers to establish the foundations of the field. Practical implementations and frameworks together with forward-thinking perspective research gave rise to the rapid growth of the field in the 2000s and 2010s, with a need for and a trend towards holistic approaches and exemplars. Throughout the evolution of the field, researchers have published a large number of perspective papers to challenge the status quo and outline the needs of practitioners.

RQ3 Which are the application domains of self-adaptive systems over time? After an initial focus on networking, web services have dominated self-adaptive systems as an application area for much of the field’s evolution, up until around 2015 when IoT and IaaS became the most frequent domains. From the beginning, the field has exhibited a large and diverse number of application domains, from robotics and networking to automotive and intelligence surveillance reconnaissance.