Inter national J our nal of Recongurable and Embedded Systems (IJRES) V ol. 15, No. 3, No v ember 2026, pp. 777 ∼ 795 ISSN: 2089-4864, DOI: 10.11591/ijres.v15.i3.pp777-795 ❒ 777 A r e view of intelligent task scheduling models f or f og computing in lar ge-scale inter net of things netw orks Sababath y P alanisamy 1 , Erum Ashraf 2 , Selv akumar Manickam 1 , Huda Mohammed Alsay ednoor 3 , Mahmood A. Al-Shar eeda 4,5 , Shankar Karuppayah 1 1 Cybersecurity Research Centre (CYRES), Uni v ersiti Sains Malaysia, Geor ge T o wn, Malaysia 2 F aculty of Engineering Sciences and T echnology , Bahria Uni v ersity , Islamabad, P akistan 3 Shatt Al-Arab Uni v ersity Colle ge, Basra, Iraq 4 Department of Computer Engineering T echnologies and Articial Intelligence, Basrah Engineering Polytechnic Colle ge, Southern T echnical Uni v ersity , Basra, Iraq 5 Colle ge of Engineering, Al-A yen Uni v ersity , Thi-Qar , Iraq Article Inf o Article history: Recei v ed Apr 20, 2026 Re vised Jun 23, 2026 Accepted Jul 18, 2026 K eyw ords: F og computing Internet of things Machine learning Reinforcement learning Resource management T ask scheduling ABSTRA CT The rapid gro wth of internet of things (IoT) applications has generated an ur - gent demand for ef cient task scheduling mechanisms that feature lo w latenc y , scalability , and resource a w areness. While fog computing can meet this demand by e xtending computing an d storage resources to the vicinity of data sources, it still confronts four core challenges: resource heterogeneity , dynamic w orkloads, mobility , and security constraints. T o address these issues, this study conducts a systematic re vie w of intelligent task scheduling for fog-enabled IoT systems. It classies the included eligible studies into v e scheduling paradigms, and inno v ati v ely proposes a cross-domain frame w ork that inte grates three types of a w areness, including priority a w areness. The subsequent w ork of this study will sort out the core strengths and limitations of e xisting rele v ant methods. Finally , open challenges and emer ging research directions, including edge arti- cial intelligence (AI), federated learning, blockchain-assisted scheduling, and hardw are-accelerated implementations are discussed. The ndings pro vide useful guidance for de v eloping scalable, secure, and adapti v e scheduling frame w orks for ne xt-generation fog-enabled IoT systems. This is an open access article under the CC BY -SA license . Corresponding A uthor: Selv akumar Manickam Cybersecurity Research Centre (CYRES), Uni v ersiti Sains Malaysia Geor ge T o wn, Pulau Pinang, 11800, Malaysia Email: Selv a@usm.my 1. INTR ODUCTION The rise of the internet of things (IoT) has seen aggressi v e deplo yment s of emer ging infrastructure comprised by billions of de vices int erconnected that generate huge v ol umes data. In man y IoT applications (e.g. smart transport ation, industrial automation, healthcare monitoring, and intelligent surv eillance), lo w latenc y , high reliability , and quality-of-service (QoS) guarantees are demanded [ 1 ]–[ 3 ]. The traditional cloud-centric computing architectures cannot satisfy such requirements because processing is centralized, netw ork access can be congested, and there are long communication delays between de vices and cloud data centers [4], [5]. F og computing has recently attracted attention as a promising paradigm to o v ercome these limitations, and to e xtend computation, storage, and netw orking resources near data sources. F og computing supports decentralized processing at the netw ork edge and intermediate fog nodes, thus reducing latenc y , relie ving J ournal homepage: http://ijr es.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
778 ❒ ISSN: 2089-4864 backhaul traf c, and impro ving conte xt a w areness [ 6 ], [ 7 ]. Ho we v er , fog en vironments are heterogeneous in nature and thus dif cult to utilize resources ef ciently . One of the k e y aspects is task scheduling, schedulers ha v e to adapt w orkloads, di v erse de vice feature sets, unpredictable netw ork en vironment, and distinct application priorities [8]. Based on the literature, at least a series of intelligent task scheduling models were proposed as it seems comple x. All the mentioned abo v e mentioned static approaches are impro v ed in performing better scheduling of the jobs using heuristic rules, metaheuri stic optimization techniques, machine learning (ML), reinforcement learning (RL), or h ybrids of them all [ 9 ], [ 10 ]. While e xisting w orks sho w a considerable performance impro v ement for dedicated cases, the di v ersity of scheduling algorithms, optimization criteria and e v aluation scenarios presents challenges in de v eloping an o v erall underst anding on their strengths and weaknesses as well as applicability to lar ge-scale IoT fog systems. Although se v eral surv e y articles ha v e re vie wed task scheduling and resource management in fog and edge computing en vironments, most e xisti ng st udies focus primarily on algorithmic cate gorization or specic optimization objecti v es such as latenc y , ener gy ef cienc y , or resource utilization. More recently , Chuan et al. [ 11 ] re vie wed the challenges and opportunities associated with task scheduling in fog computing en vironments. Jamil et al. [ 12 ] presented a comprehensi v e taxonomy of resource allocation and tas k scheduling techniques in fog computing and internet of e v erything (IoE) en vironments. Hosseinzadeh et al. [ 13 ] conducted a systematic surv e y of task scheduling mechanisms in fog computing, cate gorizing e xisting solutions according to scheduling objecti v es, optimization techniques, and system architectures. T able 1 compares the proposed surv e y with representati v e recent re vie w articles. T able 1. Comparison of the proposed surv e y with e xisting re vie w studies on fog computing task scheduling Surv e y Heuristic ML RL Hybrid Security Locality Priority Future directions Chuan et al. [11] ✓ ✓ ✓ ✓ P artial × × ✓ Jamil et al. [12] ✓ ✓ × × × × × ✓ Hosseinzadeh et al. [13] ✓ ✓ ✓ × P artial × × ✓ This surv e y ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Existing re vie w papers focused on fog and edge scheduling currently suf f er from tw o types of limi- tations: rst, the y e xhibit a supercial positioning bias, and second, the y ha v e core capability shortcomings. W i th the emer gence of ne w trends in related elds, such as learning-based and h ybrid schedulers, it has become highly necessary to de v elop a ne w re vie w that meets current de v elopment needs. This study conducts a systematic re vie w of intelligent task scheduling models for lar ge-scale IoT netw orks in fog computing en vironments. W e classify and or g anize e xisting methods, dissect their core designs, and complete a comprehensi v e comparison of all re vie wed models. W e further discuss cross-cutting intelligence and open challenges in the eld, and outline clear directions for the future de v elopment of scalable, secure, and practical scheduling frame w orks. The rest of this paper is struct ured as follo ws. Section 2 pro vides research method in this paper . F og computing and scheduling basics are summarized in section 3. T axonomy of intelligent task scheduling models is discussed in section 4. W e no w present priority-a w are, security-a w are, and locality-a w are scheduling concerns in the follo wing sections. Section 5 sho ws priority-a w are scheduling in fog computing. That is included in section 6, follo wed by open challenges and future research directions in section 7. Finally , section 8 concludes the paper . 2. RESEARCH METHOD T o ensure transparenc y , reproducibility , and scientic rigor , this re vie w adopts the structured literature re vie w method and sorts out rele v ant research on lar ge-scale IoT intelligent task scheduling in fog computing en vironments in strict accordance with four standardized phases. 2.1. Literatur e sear ch strategy T o collect rele v ant literature in the eld of fog computing task scheduling, this study conducted searches across v e authoritati v e databases, including IEEE Xplore. W e only included peer -re vie wed journal articles and conference papers, limited included publications to the period from 2020 to 2025, and carried out all searches using a core search query and its v ariants. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 777–795 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 779 ( “F og Computing” OR “Edge Computing” ) AND ( “T ask Scheduling” OR “Resource Allocation” ) AND ( “Internet of Things” OR “IoT” ) AND ( “Machine Learning” OR “Reinforcement Learning” ) (1) Additional manual searches were performed using references from highly cited surv e y papers a nd recent re vie w articles to ensure comprehensi v e co v erage of the literature. 2.2. Inclusion and exclusion criteria T o maintain the quality and rele v ance of the re vie wed studies, predened inclusion and e xclus ion criteria were applied. a. Inclusion criteria: studies were included if the y satised the follo wing conditions: - Published in peer -re vie wed journals or conference proceedings. - Written in English. - F ocused on task scheduling, resource allocation, or w orkload management in fog or edge computing en vironments. - Addressed IoT -enabled applications or lar ge-scale distrib uted systems. - Proposed, analyzed, or e v aluated intelligent scheduling approaches, including heuristic, met aheuristic, ML, RL, or h ybrid methods. b . Exclusion criteria: studies were e xcluded if the y: - F ocused e xclusi v ely on traditional cloud computing without fog or edge components. - Did not address scheduling-related problems. - W ere duplicate publications. - W ere editorials, tutorials, book chapters, theses, or non-peer -re vie wed reports. - Lack ed suf cient technical details or e v aluation results. 2.3. Study selection pr ocess This study conducted a systematic re vie w . After completing database searches, duplicate remo v al, preliminary screening, and full-te xt assessment, we included all eligible li terature, and b uilt an analytical frame w ork centered on core elements such as scheduling paradigms and optimization objecti v es. Figure 1 illustrates the o v erall literature selection process. The literature re vie w process adopted in this study is di vided into four core stages: identication, screening, eligibility assessment, and nal inclusion. All deduplication operations conducted during data collection in the identication link of the rst stage, as well as all quantitati v e data, ha v e been fully retained to guarantee the replicability of this study . This study is currently in the literature screening stage. W e ha v e screened the titles and abstracts of 421 original studies , and e xcluded 287 records that do not co v er the four core research areas (including task scheduling) in accordance with pre-set criteria. F ollo wing the initi al screening process, this study conducted eligibility assessments of the full te xts of 134 articles, and e xcluded 56 articles that did not meet the inclusion criteria; the reasons for these e xclusions are di vided into four cate gories. This re vie w ultimately included 78 studies that met all screening criteria. All studies were cate gorized by scheduling paradigm, optimization o bj ecti v e, intelligent dimension, scalability characteristic, and safety consideration. The dataset generated from this process supports the subsequent construction of the classication system, comparati v e analysis, and identication of research g aps. 2.4. Classication framew ork The selected studies were systematically classied according to the follo wing dimensions: - Scheduling paradigm (heuristic, metaheuristic, ML, RL, and h ybrid approaches). - Adaptability (static or dynamic scheduling). - Optimization objecti v es (latenc y , ener gy ef cienc y , scalability , cost, and reliability). - Intelligence dimensions (priority-a w areness, security-a w areness, and locality-a w areness). A r e vie w of intellig ent task sc heduling models for fo g computing in lar g e-scale ... (Sababathy P alanisamy) Evaluation Warning : The document was created with Spire.PDF for Python.
780 ❒ ISSN: 2089-4864 - Computational comple xity and real-time suitability . This classication frame w ork enables a comprehensi v e comparison of e xisting intelligent scheduling models while identifying research g aps and future opportunities for fog-enabled IoT systems. Figure 1. PRISMA-based literature selection process used in this surv e y 3. FOG COMPUTING AND SCHEDULING FUND AMENT ALS F og computing has g ained an increasing attention as the central architectural paradigm to o v ercome the dra wbacks of cloud-centric processing for massi v e IoT systems. W ith computations, storage, and netw orking at or near the data sources, fog computing pro vides lo w-latenc y processing (delay-sensiti vi ty), higher scale ph ysical de vices able to mee t user demand for scale up/do wn resources as needed and increased conte xt a w areness. This section briey re vie ws fog computing basics and presents the importance of task scheduling as a fundamental tool for optimal resource management in fog-dependent IoT model. 3.1. F og computing ar chitectur e in inter net of things As sho wn in Figure 2, o v erall structure of a con v entional fog computing system consists of t hree logical layers: de vice, fog, and cloud. The de vice layer in v olv es IoT sensors, actuators, and smart de vices that produce data and e x ecute computational operations. These de vices ha v e limited resource a v ailability and w ould require aid from neighboring fog nodes for processing support. The fog is composed of indi vidual fog nodes that include g ate w ays, routers, micro data centers, and edge serv ers equipped with intermediate computation and storage capabilities. W ith the fog layer performing computation near data sources, communic ation delay , and netw ork traf c are alle viated. The cloud tier pro vides centralized processing and computational storage for computation-hea vy w orkloads and w orldwide analytics that are not latenc y-sensiti v e. Unlik e edge computing, fog computing enables multi-hop processing and coordination o v er multiple intermediate nodes rather than simply at the netw ork edge. This layered structure enables fog systems to accommodate a wide-range of IoT applications that ha v e dif f erent latenc y , bandwidth, and reliability constraints. 3.2. T ask scheduling in f og computing T ask scheduling in fog computing is the act of allocating computational tasks to suitable e x ecutor nodes o v er the fog and cloud layer , as sho wn in Figure 3. Unlik e cloud scheduling, fog scheduling is subject to Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 777–795 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 781 highly dynamic en vironments such as dynamically v arying w orkloads, di v erse capabilities among nodes, v olatile netw ork connection and mobility of de vices and users. Ef cient scheduling decisions ha v e a direct ef fect on system performance characteristics, response time throughput, ener gy consumption, and QoS [12], [14]. Figure 2. Three-layer fog computing architecture for IoT en vironments Figure 3. Schematic representation of task scheduling in fog computing When scheduling tasks in the fog, schedulers usually tak e a number of f actors into account such as task ur genc y , a v ailability of resources and netw ork latenc y and e x ecution cost. According to application demands, tasks may be processed on fog nodes, migrated among dif ferent fog re gions and of oaded into the cloud [ 15 ], [ 16 ]. The inef cient scheduling can cause high latenc y , lo w resource utilization and e v en the service de gradations especially in lar ge-scale IoT [17], [18]. 3.3. Scheduling objecti v es and perf ormance metrics The aim of fog scheduling is to meet the requirements of task deadline with ef fecti v e resource utilization in distrib uted en vironment [ 12 ]. T ypical scheduling objecti v es include t he minimization of end-to-end latenc y , task w aiting and completion time as well as better load balancing between fog nodes and ener gy consumption [ 19 ]. Furthermore, the contemporary fog schedulers are increasingly in v olving f airness, reliability and security- a w areness objecti v es into their models. Authors commonly use performance measures including response time, mak espan, resource utilization, ener gy c onsumption, le v el of satisf action in task e x ecution and scheduler o v erhead to access the scheduling ef cienc y [ 20 ]. In a macro-IoT system, these schedulers also need to be e v aluated in terms of scalability and adaptability since the y should re main ef cient when task arri v al rates are l ar ge and the communication en vironment is dynamic [21]. A r e vie w of intellig ent task sc heduling models for fo g computing in lar g e-scale ... (Sababathy P alanisamy) Evaluation Warning : The document was created with Spire.PDF for Python.
782 ❒ ISSN: 2089-4864 3.4. Difculties of f og-scheduling f or massi v e inter net of things S c h e d u l i n g i n f o g c o m p u t i n g e n c o u n t e r s v a r i o u s i n t r i n s i c c h a l l e n g e s , b e c a u s e o f i t s d e c e n t r a l i z e d a n d h e t e r o g e n e o u s c h a r a c t e r i s t i c s . R e s o u r c e a v a i l a b i l i t y a n d f o g n o d e h e t e r o g e n e i t y m a k e t h e c h a l l e n g e o f p l a c i n g t a s k s i n c r e a s i n g l y d i f f i c u l t , w h i l e t h e d y n a m i c n a t u r e o f n e t w o r k i n g c o n d i t i o n s a n d n o d e s ’ m o b i l i t y b r i n g u n c e r t a i n t y a n d i n s t a b i l i t y [ 1 3 ] , [ 2 2 ] . L a t e n c y - c o n s t r a i n t a n d m i s s i o n - c r i t i c a l I o T a p p l i c a t i o n s r e q u i r e t h e p e r f o r m a n c e t o b e e n f o r c e d w i t h i n s t r i c t t i m e b o u n d s , w h i c h i s h a r d t o a c h i e v e i n a c o n s i s t e n t m a n n e r . B e s i d e s , t h e o v e r h e a d c a u s e d b y s e c u r i t y a n d p r i v a c y c o n s t r a i n t s c a n a l s o b e d e t r i m e n t a l t o s c h e d u l i n g p e r f o r m a n c e [ 2 3 ] . These challenges naturally moti v ate the design of intelligent scheduling models, which respond to v ariable conditions, trade of f competing objecti v es and act ef ciently at internet scale. In the rest of this paper , such intelligent task scheduling models are re vie wed and analyzed, their adv antages/disadv antages analyzed and applicability in lar ge-scale IoT fog computing system. 4. T AXONOMY DIMENSIONS AND CLASSIFICA TION CRITERIA F og computing has been proposed as a fundamental architect u r al paradigm to o v ercome the deciencies of cloud-centric processing in massi v e internet of things (mIoT) deplo yments. F og computing achie v es locally computationally intensi v e operations, data storage and netw orking near data sources to process at lo w latenc y , gro w in scalability and augment conte xt a w areness. This section presents fundamental concepts of fog computing and the notion that task scheduling is identied as a core technique to ef ciently manage resources in fog-oriented IoT -based en vironments. W e tak e the follo wing dimensions to consistently cate gorize the literature. - Decision mechanism: deterministic rule sets, search-based or gradient optimization based, inference of beha vior from observ ed data (data-dri v en), and polic y learning. - Adaptability: static policies vs. dynamic policies based on w orkload and the v ariation of mobility and resources. - Optimization objecti v es: single-objecti v e (e x. latenc y) vs. multi-objecti v e (vs. trade of between latenc y and ener gy and cost). - System kno wledge: information re g arding netw ork locality , re gional constraints, security o v erhead, and system di v ersity . - Computation o v erhead: whether scheduling in real-time is feasible due to resource scarcity of the fog. According to these dimensions, intelligent task scheduling solutions can be classied into v e primary types: heuristic and rule-based schemes, metaheuristic schemes, ML mechanisms, RL de vice, h ybrid or adapti v e mechanism. 4.1. Heuristic and rule-based scheduling models Heuristic and rule-based scheduling models ha v e been the rst and most popular cate gory of task schedulers in fog computing en vironments for their simplici ty , less computation, and easy implementation [ 24 ], [ 25 ]. Note that these models either based on predened rules, static priorities or light-weight decision functions used to assign tasks into a v ailable fog nodes, as sho wn in Figure 4. T ypical scheduling algorithms include, b ut are not limited to, First-In First -Out (FIFO), Shortest Job First (SJF), Earliest Deadline First (EDF), Round Robin (RR), and static priority-based schedulers [ 26 ]. In fog-based of IoT systems, such mechanism may be enriched with v ery basic conte xt data, lik e node a v ailability , queue length or distance to the data source in order to minimize latenc y and communication o v erhead [27], [28]. The adv antage of heuristic and rule-based schedulers is their speed in decision-making as the timely scheduling in resource-limited fog nodes for latenc y-sensiti v e applications a re crucial [ 29 ]. Its deterministic properties pro vide predictable system beha viour and si mple inte gration with container orchestration platforms and compact fog middle w are. Therefore, these schedulers are often tak en as basic polices or b uilding blocks i n other late schedulings [30]. Ho we v er , heuristics and rule-based models ha v e poor performance in lar ge-scale and dynamic IoT netw orks. Static logic limit s all adaptations to w orkload changes, mobile nodes, and the heterogeneous resources. Another issue with priority-based heuristics is that the y naturally lead to task starv ation or can be dened as lo w-priority tasks starving for servicing time in a high prioritization en vironment. These models al so f ail to handle multi -objecti v e optimization lik e minimizing latenc y , ener gy , and security o v erhead in parallel. Although heuristic and rule-based scheduling models are attracti v e in simple or resource constrained en vironments, the y f all short under the comple xity and scale requirements of today fogs paradigms. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 777–795 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 783 Figure 4. Ov ervie w of heuristic and rule-based task scheduling models in fog computing 4.2. Metaheuristic scheduling models D u e c o m p l e x i t y a n d m u l t i - o b j e c t i v e n a t u r e o f t a s k s c h e d u l i n g i n l a r g e - s c a l e I o T e n v i r o n m e n t s , m e t a h e u r i s - t i c s s c h e d u l i n g m o d e l s h a v e b e e n h e a v i l y a p p l i e d i n f o g c o m p u t i n g . U n l i k e h e u r i s t i c - b a s e d s o l u t i o n s t h a t u s e f i x e d r u l e s , m e t a h e u r i s t i c m o d e l s e m p l o y p o p u l a t i o n o r e v o l u t i o n a r y s e a r c h s t r a t e g i e s t o e x t e n d i n t o a w i d e r s o l u t i o n s p a c e a n d p r o v i d e c l o s e - t o - o p t i m a l s c h e d u l i n g d e c i s i o n s , a s s h o w n i n F i g u r e 5 . T h i s c a n o f t e n b e a c h i e v e d b y m e a n s o f m e t a h e u r i s t i c a l g o r i t h m s l i k e g e n e t i c a l g o r i t h m s ( G A ) , p a r t i c l e s w a r m o p t i m i z a t i o n ( P S O ) , a n t c o l o n y o p t i m i z a t i o n ( A C O ) , d i f f e r e n t i a l e v o l u t i o n ( D E ) , g r e y w o l f o p t i m i z a t i o n ( G W O ) , a n d o t h e r b i o - i n s p i r e d m e t h o d s . Figure 5. Conceptual o v ervie w of metaheuristic scheduling models in fog computing In fog systems, metaheuristic schedulers are mainly de v eloped in order to decide multiple and con- icting objecti v es (latenc y and completion time task minimization v ersus ener gy consumption reduction or w orkload balance between dif ferent fog nodes or operational sa vings). The y are e xible enough to include a v ariety of constraints such as node heterogeneity , limited netw ork bandwidth, mobility , and task deadlines. Hence, metaheuristic approaches ha v e generally pro vided better solutions than classical heuristics in comple x scheduling problems, especially when global optimization is in v olv ed. Despite their ef fecti v eness, h o we v er , metaheuristc scheduling models are certainly f aced with major obstacles in realistic fog scenarios. This tak es a higher computational cost due to the iterati v e search and e v alua- tion approach that is based on population which is impractical for applications with tight latenc y requirements or fog nodes that ha v e limited capabilities. Performance is al so susceptible to the parameterization, notably the population size, mutation rate or con v er gence thresholds, that usually ha v e to be e xtensi v ely hand-tuned. As a result, although metaheuristic scheduling optimization models pro vide po werf ul solutions to the comple x fog A r e vie w of intellig ent task sc heduling models for fo g computing in lar g e-scale ... (Sababathy P alanisamy) Evaluation Warning : The document was created with Spire.PDF for Python.
784 ❒ ISSN: 2089-4864 systems optimization problems, the y may need to be simplied or h ybridized with simpler heuristics or e v en of oaded to more capable de vices in order ensure the scalability and responsi v eness of actual deplo yments. 4.3. Machine lear ning-based scheduling models F og computing has paid considerable attention to ML-based scheduling models , in order to address the limitation of static and rule-based schedulers in lar ge-scale IoT systems, in Figure 6. No w adays the ab undance of ne grain system acti vity data can be e xploited by learning models to infer scheduling decisions from historic system beha vi or , which can help achie v e proac ti v e and adapti v e task placements under dynamic w orkload and netw ork f a ctors [ 31 ], [ 32 ]. Contrary to heuristics or metaheuristics depending on pre-selected rules or iterati v e search, ML-based schedulers predict system’ s rele v ant parameters and act based on this e xpectation for scheduling [33]. Figure 6. Conceptual o v ervie w of ML-based scheduling models in fog computing T ask e x ecution time, netw ork delay , ener gy consumption, or node a v ailability are often estimated based on past observ ations using supervised learning methods. Once ha ving the ranks, we construct re gression and classication models to rank candidate fog nodes or choose e x ecution places ag ainst predened objecti v es. Unsupervised learning techniques (e.g., clustering) are usua lly incorporated to cluster the fog nodes by resource similarity , geographic proximity , or w orkload patterns with a goal of reducing decision comple xity i n addition to enhancing scalability . Ensemble learning and deep learning methods are also proposed to model the non-linear relationships between system states and scheduling results in some studies. The pri mary adv antage of ML-based scheduling models is generalization from data, supporting predicti v e conte xt-a w are decision making. This allo ws better handling of uncertainty , w orkload di v ersication and heterogeneity among fog nodes. But the y are only as good as the data on which the y train. There may also be intensi v e computation o v erhead associated with model training and updates for resource-limited fog en vironments. In addition, performance can deteriorate because of concept shift, brought about when the dri ving conditions are f ar from the ones encountered when training data w as collected. Consequently , although ML scheduling models poss ess great e xibility and prediction ability , designing light-weight models, retraining model periodically or collaborating with simpler scheduling schemes must be emplo yed to guarantee the rob ustness and real-time feasibility . 4.4. Reinf or cement lear ning scheduling models RL [ 34 ], [ 35 ] base d scheduling models ha v e been identied as a potential candidate due to their superior intelligence c ompared to rule-based schedulers for fog computing systems, which are commonly found in lar ge-scale IoT netw orks that e xhibit dynam ic w orkloads and resources, and uctuating netw ork states, as sho wn in Figure 7. In contrast to ML-based supervised approaches which are based on l abeled histori cal data, an RL-based scheduler learns optimal scheduling policies by interacting directly with the en vironment observing system states, taking actions and recei ving feedback in terms of re w ards. In fog computing, RL has been widely emplo yed to formulate the task scheduling as a sequentia l decision making problem. What is more, state s pace can consist of; i) fog nodes utilisation, i i) queue lengths, iii) netw ork delay , and i v) task ur genc y or ener gy amount and actions may mean choosing e x ecution/of oading Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 777–795 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 785 destinations. T raditional RL algorithms lik e Q-learning ha v e been used in the conte xts of lo w-dimensional cases, while deep reinforcement learning (DRL) approaches (e.g., deep Q-netw orks (DQN), proximal polic y optimization (PPO), and actor–critic models) are being used more often to address high-dimensional state and action spaces. Figure 7. Conceptual o v ervie w of RL-based scheduling models in fog computing One of most attracti v e characteristics when dealing with RL-based scheduling approach is its high v ersatility and ef fecti v eness in terms of long-run optimization. RL schedulers can adapt by continually learning ne w policies to react to w orkload v ariation, node mo v ement, and netw ork condition changes. This mak es it suitable for real-time or mission-critical IoT applications in which static or predicti v e models may not generalize. Y et, RL-based scheduling method also comes with v arious dif culties. T raining and con v er genc e could be slo w , particularly in lar ge-scale fog en vironment with complicated state spaces. Learning-based e xploration may infer suboptimal or unsafe scheduling policies, which w ould be undesirable for real-time critical safety-related tasks. Furthermore, the computati onal cost to train deep models could be be yond the capacity of resource-limited fog nodes. As a result, practical RL-based schedulers often rely on centralized or hierarchical learning re w ard shaping, and/or with more nai v e heuristic components to achie v e stability , scalability and applicability in real-w orld scenarios. 4.5. Hybrid and adapti v e scheduling models Hybrid and adapti v e scheduling mechanisms are part of a ne w gener ation of smart schedulers de v eloped to o v ercome t h e disadv antages of single-paradigm model in fog computing, as sho wn in Figure 8. These frame- w orks combi ne a v ariety of schedul ing methods (e.g., heuristics, metaheuristics, ML, and RL) to compromise among the decision quality , adaptability , and computational ef fort in the conte xt of lar ge- scale IoT systems. Hybrid methods are primarily born out of the moti v ation to benet from conjunction of tw o approaches with complementary adv antages and counterpart disadv antage. In the fog computing domain, heterogeneity scheduling often combines lightweight heuristics with learning-based or opt imization-dri v en methods. F or e xample, initial task ltering or f ast decision making can be achie v ed by heurist ic rules and placement decisions are further rened for optimization by metaheuristic or learning-based modules. In other related w orks, M L models are used for parameter prediction in the system-side and RL is performed to dynamically adjust scheduling policies o v er time. Adapti v e scheduling models pro vide greater inte gration by inte grating feedback-based, threshold-based triggers, or periodic polic y updates that enable the scheduler to respond to w orkload v ariability , node ph ysical motion and netw ork dynamics. The k e y benet of the h ybrid and adapti v e scheduling is rob ustness and adaptability . When scheduling beha viour is dynamic and adapted to real-time system conditions, these models can result in better latenc y , resource usage, f airness and scalability than with static or single-method schedulers. This renders them well- suited for heterogeneous and non-stationary fog netw orks that serv e mission-critical IoT applications. A r e vie w of intellig ent task sc heduling models for fo g computing in lar g e-scale ... (Sababathy P alanisamy) Evaluation Warning : The document was created with Spire.PDF for Python.
786 ❒ ISSN: 2089-4864 Figure 8. Illustration of h ybrid and adapti v e task scheduling approaches in fog computing Ho we v er , the design comple xity of h ybrid and adapti v e models has become signicantly con v oluted. In order to inte grate multiple scheduler components, an in-depth coordination, parameter optimisation and picture for consistenc y management are necessary in order to pre v ent interference between decision making. Furthermore, it i s lik ely that adapti v e procedures ma y lead to control o v erhead and stability i ssues if the y are not properly designed. Therefore, although h ybrid and adapti v e scheduling schemes are seen as a promising trend in the future fog systems, their realistic deplo yment requires well consideration on architectural design and thorough e v aluation to guarantee real-time feasibility and system reliability . 4.6. Cr oss-cutting scheduling intelligence framew ork R e c e n t y e a r s h a v e w i t n e s s e d t h e l a t e s t r e s e a r c h p r o g r e s s i n t h e f i e l d o f f o g c o m p u t i n g s c h e d u l i n g , w h i c h s h o w s t h a t s c h e d u l i n g i n t e l l i g e n c e i s n o l o n g e r d e t e r m i n e d s o l e l y b y a s i n g l e u n d e r l y i n g o p t i m i z a t i o n a l g o r i t h m . E f f e c t i v e d e c i s i o n - m a k i n g m u s t r e l y o n c r o s s - p a r a d i g m , i n t e r d i s c i p l i n a r y i n t e l l i g e n c e d i m e n s i o n s , w h i c h c a n c o v e r a l l m a i n s t r e a m s c h e d u l e r s a c r o s s a l l m a j o r c a t e g o r i e s : h e u r i s t i c , m e t a - h e u r i s t i c , M L , R L , a n d h y b r i d a p p r o a c h e s . This paper combed through published re vie w literature, and e xtracted three cross-domain intelligent dimensions: priority , safety , and location a w areness, to supplement the four optimization goals of traditional technologies. T o formalize these observ ations, this surv e y proposes the cross-cutting scheduling intelligence frame w ork illustrated in Figure 9. The frame w ork inte grates multiple intelligence dimensions into a unied scheduling decision process for fog-enabled IoT en vironments. 4.6.1. Priority awar eness Priority-a w are is an important aspects of intelligence, especially for fog scheduling in IoT syst ems with mix ed timing-sensiti v e and QoS-sens iti v e applications. A number of scheduling systems ha v e introduced task priority schemes to distinguish between delay-sensiti v e, safety-critical, and best-ef fort w orkloads. Most of these priority-adapti v e schedulers preempt when high-priority tasks are ready to be run, enforce processing deadlines for jobs, or use weighted queues to ensure that high-importance w ork is performed without undue delay . Ne v ertheless, if priorities are not managed properly , lo w-priority tasks might be starv ed of resources—indicating the requirement to f airness-a w are or adapti v e priority adjustments for lar ge-scale deplo yments. 4.6.2. Security awar eness Although re g arding fog in general as the e xtension and decentralization of cloud computing which is still vulnerable to man y threats such as much high e x ecution tasks and IoT c apture enriched other , in recent years security a w areness for this setting has become more serious. Scheduling will not tak e into account security o v erhead introduced for e xample by encryption, authentication, trusted e x ecution en vironment or isolation mechanisms. Security-conscious schedulers attempt to optimize performance goals while respecting protection needs through dynam ic node selection (based on security capabilities) or adjusting le v els of security per task sensiti vity . Ho we v er , with increasing popularity there is still li mited attention in the e xisting scheduling models Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 777–795 Evaluation Warning : The document was created with Spire.PDF for Python.