PaperlayerView exact revision on arXiv

Source-linked AI summary

Autonomy and Intelligence in the Computing Continuum: Challenges, Enablers, and Future Directions for Orchestration

Henna Kokkonen, Lauri Lovén, Naser Hossein Motlagh, Abhishek Kumar, Juha Partala, Tri Nguyen, Víctor Casamayor Pujol, Panos Kostakos, Teemu Leppänen, Alfonso González-Gil, Ester Sola, Iñigo Angulo, Madhusanka Liyanage, Mehdi Bennis, Sasu Tarkoma, Schahram Dustdar, Susanna Pirttikangas, Jukka Riekki

arXiv:2205.01423v3cs.MAcs.AIcs.LG

TL;DR

The paper frames continuum orchestration as needing a balance between local autonomy and centralized control, since autonomous agents may fail to achieve system goals. It surveys edge AI methods and discusses their potential uses in orchestration functions.

  • Problem

    Local autonomy alone may not achieve system goals because nonlinear interactions make system-level behavior difficult to predict from individual agents.

    Source: B. Weakly coupled, autonomous control · p. 15

  • Method

    The paper surveys distributed machine learning, decision-making, and negotiation methods, assessing their strengths and weaknesses for computing-continuum challenges.

    Source: §VIII · p. 39

  • Results

    The survey identifies potential uses for edge AI capabilities in implementing continuum orchestration functions.

    Source: §VIII · p. 39

  • Takeaways & Limitations

    Edge AI methods may contribute to implementing the functionality required of continuum orchestration.

    Source: §VIII · p. 39

Abstract

from arXiv · show

Future AI applications require performance, reliability and privacy that the existing, cloud-dependant system architectures cannot provide. In this article, we study orchestration in the device-edge-cloud continuum, and focus on edge AI for resource orchestration. We claim that to support the constantly growing requirements of intelligent applications in the device-edge-cloud computing continuum, resource orchestration needs to embrace edge AI and emphasize local autonomy and intelligence. To justify the claim, we provide a general definition for continuum orchestration, and look at how current and emerging orchestration paradigms are suitable for the computing continuum. We describe certain major emerging research themes that may affect future orchestration, and provide an early vision of an orchestration paradigm that embraces those research themes. Finally, we survey current key edge AI methods and look at how they may contribute into fulfilling the vision of future continuum orchestration.

I. INTRODUCTION

Cloud-centric architectures fall short of future AI applications’ performance, reliability, and privacy needs, while edge deployment offers privacy and latency benefits but creates resource and coordination challenges.

I. INTRODUCTION

  • Continuum orchestration must move beyond cloud-developed centralized, reactive, threshold-based approaches toward autonomy and distributed intelligence.Source: §I · p. 1
  • Cloud-bound sensor analytics burdens networks and energy, increases processing latency, and raises privacy concerns.The paper describes bringing intelligence closer to data-generating devices, users, and controlled devices as the response.Source: §I · p. 1
  • Edge processing can improve privacy and latency, but heterogeneous nodes, intermittent links, constrained resources, and distributed non-IID data complicate AI deployment.The paper says meeting performance, reliability, and security guarantees requires new architectures and algorithms for training, inference, and decision making.Source: §I · p. 1
  • Emerging services such as computation offloading, multisite collaboration, and context awareness make automated deployment and management important for efficient resource use and performance.Source: §I · p. 2
  • The 6G vision and use cases call for autonomous, context-aware, self-learning, self-optimizing computational, communication, and storage services that support security, efficiency, and reliability.Source: §I · p. 2
  • AI-based orchestration manages, allocates, and distributes continuum resources dynamically, efficiently, and context-aware.Source: §I · p. 2

A. Contributions

The survey argues that continuum orchestration should embrace edge AI and local autonomy, and sets out a holistic view, an early orchestration vision, and research directions toward that goal.

A. Contributions

  • The survey adopts a multi-agent view of the edge as autonomous, adaptive, context-aware nodes that monitor conditions, forecast events, and react proactively while balancing multiple objectives.It envisions decentralized orchestration of computational, network, energy, and memory resources in response to dynamic conditions.Source: A. Contributions · p. 2
  • It proposes a broader definition of continuum resources and orchestration, using Everything as a Resource to simplify the overall view.Source: A. Contributions · p. 3
  • It combines existing orchestration taxonomies from the literature into a more holistic taxonomy.Source: A. Contributions · p. 3
  • It reviews current and emerging orchestration paradigms and identifies deficiencies for autonomous, intelligently optimized continuum orchestration.Source: A. Contributions · p. 3
  • It presents an early vision of autonomous, weakly coupled, and secure orchestration across the computing continuum.Source: A. Contributions · p. 3
  • It identifies opportunities, research directions, and challenges for realizing the vision, alongside an overview of edge AI methods considered key to autonomous orchestration.Source: A. Contributions · p. 3

A. Edge AI

The surveyed edge AI literature has focused more on running AI on edge infrastructure than using AI to optimize edge systems, and existing AI-for-edge surveys cover bounded methods or domains.

A. Edge AI

  • Most edge AI surveys emphasize AI on edge, including Park et al.’s survey of communication-efficient edge model building [16].Source: A. Edge AI · p. 3
  • AI for edge has received less attention; Deng et al. [11] organize it by topology, content, and service but cover AI methods for only selected problems.Source: A. Edge AI · p. 3
  • Lim et al. focus on federated learning for edge network problems, while Pham et al. provide a concise machine-learning overview within a broader MEC survey.Source: A. Edge AI · p. 3
  • Wang et al. [18] survey deep learning across edge computing, including edge management, whereas Xu et al. restrict their survey to IoV server placement and offloading.Source: A. Edge AI · p. 3

B. Orchestration

Orchestration surveys commonly cover a specific domain or part of the device-edge-cloud continuum, leaving its aspects without a holistic survey. The section reviews resource-management taxonomies and their lower-level scope.

B. Orchestration

  • Most orchestration surveys focus on a specific domain or continuum segment; the authors report no survey that holistically connects orchestration across the whole device-edge-cloud continuum.Examples of domains include containers, networks, and tasks.Source: B. Orchestration · p. 3
  • Toczé and Nadjm-Tehrani’s resource-management taxonomy groups resources by type, objective, location, and use.It lists six resource types and five management objectives.Source: B. Orchestration · pp. 3–4
  • Hong and Varghese’s fog/edge resource-management taxonomy covers architectures, infrastructure, and algorithms.Its architecture categories include data flow, control, and tenancy; algorithms include discovery, benchmarking, load-balancing, and placement.Source: B. Orchestration · pp. 3–4
  • Mampage et al.’s serverless resource-management taxonomy covers workload modelling, scheduling, scaling, deployment environment, workload management, and QoS goals.Workload management includes application requirements and structure, workload nature, and data locality.Source: B. Orchestration · pp. 3–4
  • The three resource-management surveys are limited to orchestration of lower-level fundamental, physical, and virtual elements.Source: B. Orchestration · p. 4

III. DEFINITIONS AND SCOPE

The definitions frame intelligent applications as agents that perceive their environment and act toward objectives, with autonomy and intelligence expressed through agent capabilities. Distributed AI is framed through multi-agent systems whose participants coordinate, adapt, and pursue individual or shared objectives.

A. Artificial Intelligence and Autonomy

  • AI agents pursue objectives using their capabilities, perceive their environment, and act through autonomous decision-making; their intelligence can involve reactivity, sociality, proactivity, and learning.Reactivity concerns responding to environmental change, sociality interaction with other agents, and proactivity taking initiative.Source: A. Artificial Intelligence and Autonomy · p. 6
  • Agents’ knowledge is encapsulated in models, which experience-based learning builds and adapts; decision algorithms use those models to select actions toward objectives.Reinforcement learning is described as a popular choice, with actor models selecting actions and critic models estimating accumulated reward.Source: A. Artificial Intelligence and Autonomy · p. 6
  • Autonomy means achieving a goal without external control while independently planning, deciding, and readjusting when obstacles arise.Source: A. Artificial Intelligence and Autonomy · p. 6

B. Distributed AI

  • Multi-agent systems distribute application functionality among agents that cooperate on individual objectives and collaborate on shared ones when they lack sufficient information or resources.Source: B. Distributed AI · p. 6
  • Agents coordinate by negotiating and sharing information or resources, while system behavior emerges from their interactions under orchestration or choreography.Open IoT multi-agent systems may reorganize in response to changes in participants or environment, and learning from experience can help them adapt toward proactive behavior.Source: B. Distributed AI · p. 6

C. Computing continuum

This section examines computing-continuum resources and orchestration, defining their scope while assessing orchestration paradigms and their limitations. It frames research themes and an early vision for orchestration, including challenges in managing emergent behavior and enabling efficient communication of shared meaning among autonomous systems.

C. Computing continuum

  • The continuum’s resources span abstraction levels from energy and time to workflows, fall into communication, computation, and data categories, and are unified by an Everything as a Resource view.Higher-level resources depend on lower-level ones, but the hierarchy is not a strict layered architecture; workflow data may come directly from cyber-physical sensors.Source: C. Computing continuum · p. 7, Figure 2 · p. 7

D. Orchestration

  • Orchestration is holistically defined as managing continuum resources from fundamental to workflow levels through functions including lifecycle management, allocation, discovery, dataflow management, and monitoring.Implementations may vary in security, privacy, fault tolerance, and control topology; objectives may reflect resource budgets and stakeholder priorities.Source: D. Orchestration · pp. 8–9

E. Orchestration control topologies

  • Centralized control leaves little autonomy to local components, while decentralized and distributed topologies increase autonomy; decentralized systems combine centralized cluster control with peer-to-peer inter-cluster coordination.Distributed scheduling can account for task deadlines and privacy, for example by offloading public, loose-deadline tasks to the cloud and keeping private, urgent tasks local.Source: E. Orchestration control topologies · p. 9

F. Scope

  • The survey focuses on decentralized and distributed control, modeling continuum systems as autonomous intelligent agents that pursue local objectives while remaining synchronized with global objectives.It assumes agent models are primarily neural networks and also addresses security and privacy implications of integrating AI into the continuum.Source: F. Scope · p. 10

IV. ORCHESTRATION PARADIGMS

  • The orchestration-paradigms chapter classifies mechanisms by computing resources in Sections IV-A–IV-C, then addresses edge- and fog-specific solutions in Sections IV-D–IV-E.Each section presents core concepts, state-of-the-art examples, and suitability for the computing continuum.Source: §IV · p. 10

A. Container orchestration

  • Cloud-native container platforms use centralized control and emphasize container lifecycle management, but are unsuitable as the continuum’s full orchestration solution given its heterogeneity and intermittent communication.They may still serve as components of a broader continuum orchestration framework.Source: A. Container orchestration · p. 11

B. Workflow management systems

  • Workflow management systems provide infrastructure for setting up, specifying, executing, and monitoring workflows composed of tasks running on available computing resources.Source: B. Workflow management systems · p. 11
  • Workflow systems manage workflow and application resources through functions such as scheduling and optimization, but do not address continuum communication, computation, and data challenges; the cited non-IID learning proposal is limited to federated learning [68].Examples include Airflow, Kubeflow, and MLflow, which support workflow execution and ML lifecycle management.Source: B. Workflow management systems · p. 12

C. Network orchestration

  • Network orchestration maps services to network resources and manages their lifecycles, but current work largely omits application and workflow needs, while vendor interoperability and centralized SDN remain concerns [26].Studies target resource utilization, VNF placement, routing, slicing, and automated service management –.Source: C. Network orchestration · pp. 12–13

D. Edge/fog orchestration

  • Edge/fog orchestration accounts for scarce cyber-physical, communication, computation, and data resources, but available approaches have gaps in control design, centralization, autonomy, and lightweight-device support.ETSI MEC is centralized; KubeEdge and OpenYurt offer edge-layer autonomy when disconnected, but containerization often excludes the lightest devices.Source: D. Edge/fog orchestration · p. 13

1) Multi-domain orchestration:

  • Multi-domain orchestration can provision resource-hungry services across providers, but raises privacy, security, heterogeneity, and incentive-alignment challenges; current network- or container-focused efforts cover only part of continuum needs.Containerization also remains unsupported on many lightweight devices without custom solutions.Source: 1) Multi-domain orchestration · pp. 13–14

2) FaaS orchestration:

  • FaaS shifts resource allocation and runtime management to providers and can place stateless functions across continuum tiers, but stateful or data-heavy workloads complicate migration and workflow-state management.FaaS can also impose vendor lock-in and make testing and debugging complex as provider infrastructure changes.Source: 2) FaaS orchestration · p. 14

3) WASM:

  • WebAssembly offers a compact, low-overhead alternative to containers for resource-poor edge devices and FaaS, but distributed WASM runtimes in the continuum lack orchestration frameworks in the authors’ assessment.Early cloud-based WASM orchestration frameworks exist, including WASMcloud [96].Source: 3) WASM · pp. 14–15

A. Computing Continuum management model

  • Continuum management must be holistic and adaptive because resources are heterogeneous and dynamic, the system is non-elastic, and stakeholders’ objectives and system balance remain difficult to measure and enforce.Research needs include flexible resource representations, tracking temporal and causal changes, proactive adaptation against failure propagation, and learning-based management.Source: A. Computing Continuum management model · p. 15

B. Weakly coupled, autonomous control

  • Continuum orchestration should balance local autonomy with minimal centralized control, which supports fair allocation, agent cooperation, and global optimization while limiting interference and resource waste.A hierarchy can use learned multi-objective optimization to allocate resources and update goals and constraints as conditions change.Source: B. Weakly coupled, autonomous control · pp. 15–16
  • In loosely coupled distributed systems, a key challenge is preventing harmful emergence while harnessing emergent activity to achieve system-level goals.Source: B. Weakly coupled, autonomous control · p. 16

C. Semantic communication in the computing continuum

  • Traditional communication theory treats message meaning as irrelevant, while interacting intelligent agents need information notions that capture shared context and learning to communicate changes in that context [103] [104].Source: C. Semantic communication in the… · p. 16
  • A proposed framing uses von Neumann–Morgenstern sequential games to treat information as imperfect state knowledge, with agents exchanging minimally sufficient information states for shared goals to reduce classical transmissions.Source: C. Semantic communication in the… · p. 16
  • The prerequisites for this efficient joint behavior and the mathematical characterization of such networked systems remain unknown.Source: C. Semantic communication in the… · p. 16
  • Open questions include when cooperative communication emerges robustly across agents with differing priors, beliefs, knowledge, or architectures; when agents converge on a shared language; and how they cope with limited knowledge, compute, and memory.Source: C. Semantic communication in the… · p. 16

D. Security and Privacy

The section presents edge AI as a way to bring security capabilities closer to users and keep more processing local, while emphasizing that decentralization and edge deployment introduce their own security, privacy, and trust challenges.

D. Security and Privacy

  • Local edge processing can reduce raw-data transfer over untrusted backhaul, support early attack detection, and limit some attacks to a local vicinity rather than a centralized single point of failure.The authors also describe local processing as reducing privacy leakages associated with long-distance transmission, while early detection may limit attack propagation to critical core-network elements.Source: D. Security and Privacy · pp. 16–17
  • Edge AI can support more adaptive Security-as-a-Service, including intelligent packet filtering and rate adjustment to help limit DDoS effects on cloud resources.The proposed examples include timing out half-open connections and dropping packets likely to be part of an attack.Source: D. Security and Privacy · pp. 16–17
  • Physical exposure makes edge AI devices vulnerable to tampering, malware, device swaps, and rogue data or algorithms, while captured AI devices may jeopardize localized services.The section names hardware roots of trust, cryptographic IDs, encryption in flight, and automated patching as possible safeguards; decentralization also brings novel attack surfaces.Source: D. Security and Privacy · p. 17
  • Keeping data at the edge does not guarantee privacy: the benefit depends on edge servers having higher privacy and trust levels than cloud servers, and multi-domain stakeholders make trust difficult to establish.Open implementations, clear specifications, and reputation systems are among the suggested trust strategies.Source: D. Security and Privacy · p. 17
  • Edge AI must address attacks on data and models, including injection, poisoning, evasion, model inversion, and membership inference; poisoning remains largely unsolved.Even a small number of poisoned federated-learning updates can substantially reduce classification accuracy and recall.Source: D. Security and Privacy · p. 17
  • Privacy and integrity protections carry costs: anonymization can reduce classification performance and convergence or struggle with images, while cryptography and secure multiparty computing can demand substantial computation or communication.The paper calls for balancing privacy protection, performance, computation, and communication costs.Source: D. Security and Privacy · p. 17

E. Synthesis

The synthesis proposes decentralized orchestration through a hierarchy of autonomous agents with objectives passed down across clusters and negotiation across administrative boundaries. It connects this vision to continuum challenges and surveys distributed learning as one potential enabler, while identifying open questions about the context, data, adaptation, and security those methods require.

E. Synthesis

  • The proposed orchestration vision models continuum resources as agents that pursue stakeholder objectives, make orchestration decisions, and may negotiate resource provisioning; objectives flow from stakeholder agents through clusters and sub-clusters.Negotiation may be cooperative within administrative domains and competitive across boundaries; the framework uses weak coupling and autonomous decisions [101].Source: E. Synthesis · p. 18, Figure 5 · p. 18
  • Distributed, heterogeneous resources, intermittent communication, siloed non-IID data, and many stakeholders with potentially conflicting objectives motivate decentralized orchestration, weak coupling, local autonomy, distributed edge intelligence, multi-objective balancing, and security and privacy protections.These requirements are summarized in Fig. 6.Source: E. Synthesis · p. 18, Figure 6 · p. 18
  • Open questions include the minimum shared context and standards agents need for interoperability, resource discovery, distributed learning, decision-making, and negotiation within and across administrative boundaries.The needed common context may depend on agents' available computation, data, and communication links.Source: E. Synthesis · p. 19
  • The synthesis also flags unresolved questions about learning convergence, scarce or siloed training data, adapting without reacting to temporary changes, and protecting data and model integrity throughout their lifecycles.Possible directions for data scarcity include pre-training, unlabeled data, and learning from few samples.Source: E. Synthesis · p. 19

VI. SURVEY OF EDGE AI APPROACHES

  • The survey introduces distributed and secure methods for machine learning, decision-making, and negotiation as promising approaches, prioritizing a map of research fields over an in-depth analysis of each method.It frames these approaches as potential responses to some of the challenges in the proposed orchestration vision.Source: §VI · p. 19

A. Distributed Learning

  • Agents need machine-learning models to predict processes and future states, but limited local data and computing resources can make accurate model training and inference difficult, motivating collaboration among nearby agents.The survey focuses this subsection on supervised learning methods that may support decentralized training and inference.Source: A. Distributed Learning · p. 19

1) Federated Learning:

The section reviews federated learning and federated distillation as distributed training approaches, describing their communication, resource, and data challenges. It also reports FedProx results under non-IID data and stragglers, and notes the communication–accuracy trade-off of exchanging model outputs instead of parameters.

1) Federated Learning:

  • Federated learning trains a global model from decentralized local data: agents periodically send local parameter updates to a central server for aggregation and receive the resulting model back.The described topology is a star; FedAvg weights local parameters when calculating the global update.Source: 1) Federated Learning · pp. 19–20, Figure 7 · p. 20
  • Vanilla FL faces communication overhead from large model updates and stragglers caused by heterogeneous hardware, energy, or connectivity; proposed responses include compression, asynchronous communication, participant selection, redundancy, and prediction of resource availability.Offloading to a neighboring node is another option, but may affect privacy, latency, and model accuracy [17].Source: 1) Federated Learning · p. 20
  • Non-IID data creates problems for consensus-model training; sharing 5% of agents’ local data significantly increased accuracy in one study, but compromises privacy and communication efficiency [138].Federated augmentation instead shares a generative model that agents use to augment data locally [139]; incentive mechanisms may reward higher-quality data and contribution.Source: 1) Federated Learning · pp. 20–21
  • With 90% of agents as stragglers, FedProx significantly improves convergence rate compared with FedAvg across five non-IID datasets.It adds a proximal term to keep local updates closer to the received global model and allows partial local work to enter the global update.Source: 1) Federated Learning · p. 21

2) Federated Distillation:

  • Federated distillation exchanges model outputs rather than parameters, reducing communication payload in proportion to output dimension and allowing agents to use different model sizes, but it is more vulnerable to non-IID data and usually less accurate than FL [139].A proposed mobile-communication approach uses FL for downlink and FD for uplink, with server-side knowledge distillation.Source: 2) Federated Distillation · p. 21

3) Fog Learning:

Fog Learning uses a layered, cluster-based architecture to move collaborative training beyond centralized star-topology federated learning. Its multi-stage hybrid algorithm aggregates within network layers and reports energy and communication savings while approaching centralized gradient-descent accuracy.

3) Fog Learning:

  • Fog Learning organizes training nodes, cluster layers, and a top-level aggregation server into a hierarchy, allowing model parameters to travel across multiple layers and clusters to change dynamically during training.Clusters may collaborate using device-to-device communication.Source: 3) Fog Learning · p. 21
  • MH-FL performs local aggregations at each network layer: D2D-enabled clusters can use distributed aggregation, while clusters without D2D use parent-node aggregation similar to conventional FL.For distributed aggregation, a parent fetches parameters from one cluster node and scales them by the cluster size to approximate the parameter sum.Source: 3) Fog Learning · pp. 21–22
  • MH-FL approaches centralized gradient-descent learning accuracy, with 50% average device energy savings and an 80% reduction in parameters transferred over network layers compared with conventional FL.Source: 3) Fog Learning · p. 22

4) Peer-to-peer Learning:

The section surveys peer-to-peer learning approaches for decentralized and self-interested agents, then considers efficient inference and the decision-making capabilities needed for autonomous continuum orchestration.

4) Peer-to-peer Learning:

  • Decentralized learning replaces central aggregation with neighborhood consensus; GADMM uses a logical chain, while dynamic variants change physical neighbors, and static primal-dual methods outperform primal methods.GADMM also limits communication to two agents in the opposite group per agent.Source: 4) Peer-to-peer Learning · p. 22
  • CTA combines neighborhood parameters before gradient updates and propagates aggregate information; ATC reverses the steps and incorporates neighbors’ gradients, while gradient exchange raises communication costs.Source: 4) Peer-to-peer Learning · pp. 22–23
  • CFA-GE combines consensus aggregation, neighbor gradients, and local SGD; its asynchronous variant predicts gradients from past intermediates to avoid waiting on neighbor exchanges.Source: 4) Peer-to-peer Learning · p. 23
  • Biscotti combines blockchain, privacy, and poisoning defenses; experiments show resilience to label-flipping attacks with 30% malicious peers, but privacy-accuracy and large-model communication costs remain.Multi-Krum validation may reject updates from peers with non-IID data, and stake-based selection does not account for peers turning malicious after gaining stake.Source: 4) Peer-to-peer Learning · pp. 23–24
  • A reputation scheme supports selective cooperation: agents in 10-agent experiments stop sharing when communication costs outweigh accuracy gains, and convergence improves over two comparison settings.The comparisons are against self-interested agents without reputation and fully cooperative agents.Source: 4) Peer-to-peer Learning · p. 24
  • Collaborative learning targets personalized local models by smoothing parameters among neighbors with similar objectives while limiting accuracy losses from smoothing.An asynchronous gossip algorithm based on ADMM solves the partial-consensus formulation; a simpler decentralized block-coordinate method is also proposed.Source: 4) Peer-to-peer Learning · p. 24

5) Efficient Model Inference:

  • Resource-constrained nodes need adapted ANN models because cloud inference raises latency and privacy concerns; compression, conditional computation, and model splitting trade off accuracy, latency, privacy, and resources.Source: 5) Efficient Model Inference · p. 24

B. Decision Making and Negotiation

  • Continuum agents need adaptive sequential decision strategies; reinforcement learning is emphasized, while offloading decisions among self-interested agents require negotiation and communication requires a shared language.Source: B. Decision Making and Negotiation · pp. 25–26

1) Reinforcement Learning:

The survey examines reinforcement-learning approaches for continuum orchestration, where agents face non-stationarity, partial observability, scalability constraints, and competing objectives. It highlights adaptations toward decentralized, privacy-preserving, and multi-policy learning, while noting open challenges for realistic deployments.

1) Reinforcement Learning:

  • Opponent-induced non-stationarity violates the stationary-environment assumption behind single-agent RL, while Nash equilibrium may be non-unique and need not maximize global reward.The survey notes that Nash equilibrium also presumes perfectly rational agents capable of infinite mutual modelling, an unrealistic assumption for resource-limited continuum agents.Source: 1) Reinforcement Learning · p. 26
  • Partial observability complicates MARL; the survey highlights agent models in MADP as a way to represent agents’ information states, capabilities, and belief updates.In fully cooperative settings, POSGs reduce to dec-POMDPs, which are NEXP-complete to solve; many proposed methods therefore rely on some centralization.Source: 1) Reinforcement Learning · p. 27
  • Methods for handling opponent non-stationarity range from forgetting old information to modelling opponents, but stronger modelling approaches can demand substantial computation, memory, and observability.For example, WoLF-PHC adapts learning rates based on whether an agent is winning or losing, while ReDVaLeR assumes observable opponent actions and repeated games.Source: 1) Reinforcement Learning · pp. 27–28

2) Deep Reinforcement Learning:

  • In mixed settings, CTDE’s dependence on centralized training information limits its fit for dynamic continuum deployment, motivating decentralized training with local, potentially intermittent communication.The survey also identifies mixed-setting MADRL as underexplored and difficult to analyze with nonlinear neural-network approximations.Source: 2) Deep Reinforcement Learning · pp. 28–29

3) Federated Reinforcement Learning:

  • Federated RL could let agents collaboratively train actor or critic networks by exchanging encrypted model parameters rather than observations, but real-life continuum applications remain sparsely studied.The survey also identifies decentralized federated learning for RL as an open research direction.Source: 3) Federated Reinforcement Learning · p. 29

4) Multi-objective Reinforcement Learning:

  • Because scalarization needs prior objective preferences and may miss non-convex Pareto-front solutions, continuum MORL should learn a spread of policies in one run.The survey identifies multi-agent learning that handles adaptation, changing preferences, and limited resources as an open question.Source: 4) Multi-objective Reinforcement… · p. 30

5) Transfer Learning:

The survey reviews research themes that may support orchestration across heterogeneous, resource-limited continuum agents. It emphasizes knowledge reuse, scalable and adaptive decision-making, bounded-cost reasoning, and negotiation, while identifying substantial open issues in real-world applicability.

5) Transfer Learning:

  • Transfer learning can reuse knowledge from cloud agents, simulations, or experts to help resource-limited agents learn, but heterogeneous agents may not share a compatible knowledge representation.The survey says current work relies heavily on human effort and is validated mainly in simple-problem simulations.Source: 5) Transfer Learning · pp. 30–31

6) Mean-field Theory:

  • Mean-field games recast an N-player game as a two-player game against a virtual average agent, but their validity for partially observable real-life MARL settings needs further study.The cited mean-field Q-learning and actor-critic approach uses the mean action of an agent’s neighborhood and requires observing neighbors’ actions and rewards.Source: 6) Mean-field Theory · p. 31

7) Evolutionary Game Theory:

  • Evolutionary algorithms can address large state spaces, partial observability under slowly changing environments, and non-stationarity, but their experience needs and risky generated policies make them unsuitable for online learning.The survey therefore points to training these algorithms in environment simulations; mutations can also send rarely visited states toward poor actions.Source: 7) Evolutionary Game Theory · p. 31

8) Swarm Intelligence:

  • Swarm intelligence is presented as a route to decentralized, adaptive coordination by resource-limited agents, whose group behavior can emerge from local interactions rather than complex individual learning.The survey discusses well-informed agents that implicitly guide group decisions and adaptive personality traits that balance individual and collective utility.Source: 8) Swarm Intelligence · pp. 31–32

9) Metareasoning:

  • Metareasoning monitors an agent’s decision process so it can make good-enough choices when time or resources do not permit thorough deliberation.Meta-level control can stop or adjust decision algorithms, choose which to run, or decide whether communication or coordination is worthwhile.Source: 9) Metareasoning · p. 32

10) Negotiation:

  • Continuum negotiation must address incomplete information, autonomy, loose coupling, and security, with mechanisms suited to agents’ differing resources and administrative settings.Opponent modelling can infer preferences from exchanged offers without requiring agents to disclose private information; suitable approaches remain an open question because evaluation benchmarks are lacking.Source: 10) Negotiation · pp. 32–33

11) Learning to Communicate:

The section examines learned agent communication and generative models as parts of an autonomous computing continuum, while identifying significant limits in their current suitability for complex settings. It highlights commitment-based protocols and generative-model capabilities alongside open challenges in communication and GAN training and evaluation.

11) Learning to Communicate:

  • RIAL and DIAL use CTDE in fully cooperative settings; DIAL passes communication-channel gradients during training, while RIAL uses Deep Recurrent Q-Networks.Both use an ANN to output values for environmental and communication actions; RIAL can also use parameter sharing.Source: 11) Learning to Communicate · p. 33
  • Learning communication protocols for large, non-cooperative or mixed agent settings remains scarce, with trust, reputation and truthful messaging still needing attention.Open settings also raise unresolved questions about how new or returning agents adapt to evolving message meanings.Source: 11) Learning to Communicate · p. 34
  • Learned protocols may require agents to share learning methods and models, and DNN-based approaches can consume too many resources for continuum agents.The section also argues that traditional protocols' fixed message flows and limited primitives can tightly couple agents and constrain autonomy [33, pp. 108-114].Source: 11) Learning to Communicate · p. 34
  • Commitment-based protocols [221, 222, 33, pp. 118-121] can define application-specific message meanings through commitments without imposing a specific message order.The interpretation can be derived from the protocol specification, supporting loose coupling and agent autonomy.Source: 11) Learning to Communicate · p. 34

1) Generative Adversarial Networks:

  • Generative models can represent high-dimensional distributions and simulate future states, letting agents evaluate actions in an imagined environment without harm from mistaken actions.They can also guide exploration and support inverse reinforcement learning [223] [225].Source: 1) Generative Adversarial Networks · p. 34
  • GANs set a generator against a discriminator in a zero-sum game; their training seeks a Nash equilibrium, but can suffer unstable training, mode collapse and vanishing gradients.Wasserstein GAN replaces Jensen-Shannon divergence with Earth Mover's distance and has shown improved training stability [228].Source: 1) Generative Adversarial Networks · pp. 34–35
  • GAN evaluation metrics remain unsettled, and most GAN research has focused on images, leaving applications to other computing-continuum problems open.Synthetic data can also augment training and mitigate non-IID data in distributed learning [139].Source: 1) Generative Adversarial Networks · pp. 34–35

2) Few-shot learning:

Few-shot learning addresses the difficulty of training in dynamic continuum settings where data is scarce and new patterns may emerge. The section focuses on meta-learning for rapid adaptation, while outlining unresolved assumptions and limits in current methods and distributed-learning work.

2) Few-shot learning:

  • Few-shot learning is relevant when continuum agents have little data and must adapt quickly to new tasks, but its methods face open issues in task relatedness, distribution shifts and distributed deployment.Meta-training assumes source tasks relate sufficiently to new tasks for adaptation to improve performance; negative transfer should be avoided.Source: 2) Few-shot learning · pp. 35, 37
  • Few-shot learning leverages prior knowledge because few samples make the empirical risk minimizer unreliable; methods use data, model or algorithm-based strategies.The article focuses on meta-learning because it supports quick adaptation across supervised and reinforcement learning problems without hand-crafting a task-specific learning algorithm.Source: 2) Few-shot learning · pp. 35–36
  • MAML learns globally useful initial parameters that can be adapted to new tasks with a few gradient steps, and applies to gradient-trained models across several learning problems.The section also describes Meta-SGD's meta-learned learning rate and MetaNAS's task-specific architecture adaptation.Source: 2) Few-shot learning · p. 36
  • Probabilistic MAML learns a distribution over global parameters, allowing multiple candidate models to be sampled before adapting to a new task.This addresses MAML's limitation of learning only a single function when few samples leave the task model ambiguous.Source: 2) Few-shot learning · p. 36
  • Few-shot methods often rely on image-classification work and assumptions about reliable labels and large labeled prior datasets; adapting to new classes also risks catastrophic forgetting [235].In a dynamic continuum, the section notes that a security model should retain previously learned attack classes while adapting to new ones.Source: 2) Few-shot learning · p. 37
  • Existing few-shot methods commonly assume task data distributions remain unchanged, while preliminary federated few-shot work can be computationally heavy and has limited experimental scope.The cited work considered supervised classification with up to 30 participants and did not account for stragglers or domain shift.Source: 2) Few-shot learning · p. 37

3) Self-supervised learning:

The section reviews self-supervised learning as a way to use unlabeled data and describes its main method families and limitations. It then maps potential edge-AI techniques to continuum orchestration functions, including monitoring, dataflow, discovery, allocation and lifecycle management.

3) Self-supervised learning:

  • Generative SSL recovers the data distribution without downstream-task assumptions, but has underperformed contrastive methods in image classification and can be sensitive to rare samples.The section links that sensitivity to reconstruction losses that model data point by point.Source: 3) Self-supervised learning · p. 37
  • Contrastive SSL assumes classification downstream tasks; its success in image classification does not establish suitability in other domains, and negative-sampling choices remain open [252].The section notes that both the role and necessity of negative sampling remain unresolved.Source: 3) Self-supervised learning · pp. 37–38
  • Adversarial SSL methods mainly rely on GANs, which require adaptations such as an encoder or partial-input reconstruction to learn representations; broader applications remain a research direction.The section notes that GAN-based approaches have mainly been used in vision applications [253].Source: 3) Self-supervised learning · p. 38
  • Self-supervised learning extracts supervision from unlabeled data to learn representations for downstream tasks, potentially reducing—but not eliminating—the labeled data needed.The section reports better performance than supervised learning when labeled data is scarce, but not when enough labels are available; it also cites robustness and out-of-distribution detection benefits.Source: 3) Self-supervised learning · pp. 37–38
  • Distributed SSL still needs evaluation with heterogeneous resource-limited participants, non-IID data and complete decentralization, while pretext-task design and model adaptation under shifts remain challenges.The section also calls for stronger theory and less ad hoc, human-dependent pretext-task design.Source: 3) Self-supervised learning · p. 38

VII. EDGE AI FOR CONTINUUM ORCHESTRATION

  • The orchestration vision treats continuum nodes as autonomous, intelligent and self-interested agents that make decisions across orchestration functions.The section presents edge AI as potentially useful across the functions summarized in Table III.Source: §VII · p. 38, Table III · p. 38
  • Predictive ML can support monitoring by forecasting changing conditions and estimating a node's status when direct information is unavailable.Within an administrative domain, monitoring agents could use collaborative peer-to-peer learning.Source: §VII · p. 38
  • Distributed learning and decision-making can inform workload placement and resource allocation, negotiation can support discovery, and decision methods can guide service lifecycle actions.Examples include deciding whether to execute or offload workloads, when to migrate containers or replicate services, negotiating resource prices across domains, and starting, stopping or updating services.Source: §VII · p. 39

VIII. CONCLUSION

The article argues that continuum resource orchestration should use edge AI to support local autonomy and intelligence. It defines continuum orchestration, surveys paradigms and research themes, and considers how edge AI methods may fulfill orchestration functions.

VIII. CONCLUSION

  • Resource orchestration in the device-edge-cloud continuum should embrace edge AI and emphasize local autonomy and intelligence.Source: §VIII · p. 39
  • The article defines continuum orchestration, assesses current and emerging paradigms, and presents an early vision shaped by emerging research themes.Source: §VIII · p. 39
  • The survey covers distributed machine learning, decision-making, and negotiation, relating their strengths and weaknesses to continuum challenges and potential orchestration functions.Source: §VIII · p. 39

Related papers

Loading 2205.01423v3…