Edge-Cloud Hybrid Task Scheduling Using Federated Reinforcement Learning and Adaptive Swarm Intelligence

The rise of edge computing can be credited to the rapid increase in technologies like Internet of Things (IoT), wireless sensor networks (WSN), mobile applications and real time data processing [1–3]. This is because edge computing allows computation of data to happen closer to the data sources resulting in better efficiency. In other words, we can say that edge computing allows data to be processed locally by the users. This local processing of data helps in minimizing time delays and improves response time. However, we cannot ignore the fact that edge resources are restricted by their computational power and storage capacity and because of these limitations it becomes difficult to do complex processing directly on the edge devices. Hybrid edge cloud architectures were developed in order to overcome these issues which allowed delay sensitive tasks to be handled by the edge nodes and compute intensive tasks were offloaded to cloud data centres [4, 5]. To optimize key performance indicators like system performance, load balancing, energy consumption QoS requirements in a distributed and heterogeneous cloud environment, intelligent task scheduling becomes an important factor [6].

Most the traditional task scheduling algorithms in cloud computing were developed to work on static workloads or centralized mechanisms. These methods often find it difficult to adapt to dynamic workloads, resource heterogeneity and latency-critical scenarios that are typically found in edge computing environments [7]. Thus, resulting in degradation of overall efficiency and response time in case of practical implementations. Moreover, centralized learning-based scheduling algorithms require continuous data aggregation, which raises concerns over privacy, bandwidth usage, and system scalability. These are critical system resources that cannot be compromised. On the other hand, bio-inspiofferslgorithms like the Artificial Bee Colony (ABC) offer adaptive capabilities but lack contextual awareness and long-term learning which limits their ability to respond effectively to evolving task patterns and resource availability [8, 9]. Hence, more intelligent and context-driven scheduling approaches need to be developed to overcome these limitations.

This paper proposes a novel task scheduling framework that addresses all these issues mentioned above. The proposed framework integrates FRL with adaptive swarm intelligence, specifically a modified ABC algorithm. FRL enables distributed edge agents to learn optimal task scheduling policies by interacting with their local environment. It also allows collaborative training a global policy without sharing raw data. This preserves data privacy, reduces network load and facilitates scalable coordination across all edge nodes [10]. This method ensures that each node improves autonomously while still benefitting from the collective intelligence. At the same time, the adaptive ABC component enhances local task placement decisions by leveraging the collective behaviour of bee-inspired agents that dynamically adjust search strategies based on feedback from the environment. The proposed algorithm can adapt to workload variations and improving resource utilization by continuously refining its search patterns.

The proposed framework is designed to optimize multiple objectives simultaneously, including task makespan, energy efficiency, resource utilization and Service Level Agreement (SLA) compliance. The proposed model combines the global learning capabilities of FRL with the adaptive search efficiency of swarm intelligence to achieve robust and intelligent scheduling in edge cloud environments.

The key contributions of this work are as follows:

• We design a Federated Reinforcement Learning-based multi-agent framework for distributed task scheduling that preserves data privacy and reduces communication overhead.

• We develop an Adaptive Artificial Bee Colony algorithm tailored for dynamic task allocation in edge-cloud environments, supporting load balancing and energy-aware scheduling.

• We formulate a multi-objective optimization problem that jointly considers task makespan, energy consumption, SLA satisfaction, and system throughput.

• We evaluate the proposed approach through extensive simulations, demonstrating superior performance over existing centralized RL and traditional swarm intelligence methods.

The rest of this paper will be structured in the following way. Section 2 outlines an overview of the literature of task scheduling, federated reinforcement learning, and bio-inspired optimization on edge-cloud. Section 3 develops Hybrid edge-cloud scheduling system model, performance objectives and constraints. Section 4 outlines the proposed Hybrid FRL-A-ABC architecture, such as the state representation, the formulation of the reward, the interaction between the FRL and the A-ABC layers, and the scheduling algorithm in general. Section 5 describes the experimental setting, the workload models, network and energy assumptions, baselines, and metrics of evaluation. The results, with their comparative analysis, ablation studies, and sensitivity analysis, are reported in Section 6. Lastly, Section 7 is the conclusion of the paper and the discussion of future research directions.

Energy-aware task scheduling for fog environments has been explored using bio-inspired heuristics such as ant-mating optimization, which reduces energy consumption but relies on a largely centralized view and offers limited adaptability to rapidly changing workloads [11]. Privacy-preserving coordination is introduced through secure federated learning for edge–cloud smart grids, while actor–criticbased Edge–AI further improves loT service provisioning in federated edge settings, though both lines of work focus more on learning and security than on rich multi-objective scheduling [12, 13]. At the edge, intelligent task allocation using machine learning and bio-inspired algorithms enhances local scheduling quality, and reinforcement-learning-based multi-objective load balancing extends this to edge-fog-cloud topologies, yet scalability and federated coordination remain open issues [14, 15]. A survey of offloading in federated cloud–edge–fog systems highlight the need to combine traditional optimization, machine learning, and federated training to handle non-stationary, heterogeneous workloads [16]. Recent studies on federated reinforcement learning for dynamic scheduling in cloud-edge-terminal networks and vehicular edge computing demonstrate improved coordination and privacy but incur significant communication cost and provide limited support for fine-grained local queue optimization [17, 18]. Pure RL-based schedulers for heterogeneous end–edge–cloud environments and unsupervised cluster Q-learning for federated edge clouds show benefits in makespan, throughput, and energy minimization, while hybrid cloud–edge deep learning frameworks scale IoT resource optimization but still treat global placement and local execution ordering as largely decoupled problems [19–21].

Taken together, existing work demonstrates (i) energy-aware heuristics at the fog/edge [11], (ii) secure FL/FRL for distributed learning and policy sharing [12, 13, 15, 18, 19], and (iii) RL-based scheduling for heterogeneous tiers [20]. However, three gaps persist: (1) limited methodological coupling between a global, privacy-preserving offloading policy and a local, adaptive queue-reordering heuristic; (2) incomplete treatment of multi-objective trade-offs (makespan, energy, SLA/tardiness, and load balance) under non-IID workloads, client sampling, and straggler effects; and (3) insufficient transparency on communication costs and convergence behavior in FRL. This work addresses these gaps by integrating FRL for where tasks are executed with an adaptive Artificial Bee Colony (ABC) module for how tasks are ordered locally, under a unified multi-objective formulation that preserves data privacy, adapts to non-IID dynamics, and reports convergence and overhead comprehensively. Table 1 presents the summary of the related work.

The proposed methodology integrates FRL with modified ABC algorithm to achieve efficient task scheduling across an edge cloud computing environment. In this hybrid framework the multiple edge nodes act as intelligent agents, each equipped with local schedulers that collaborate with a centralized cloud layer through federated updates. The workflow of this proposed work is illustrated in Figure 2.

Figure 2.

Workflow of the proposed work

Initially, each edge node observes its local state, including parameters such as CPU utilization, task queue length, energy level, task deadlines, and network latency. This can be represented as a state vector as follow in Equation (1): 1Sti=[ Cti,Qti,Eti,Dti,Lti ]S_t^i = \left[ {C_t^i,Q_t^i,E_t^i,D_t^i,L_t^i} \right]

Where:

StiS_t^i : State of edge node i at time t

CtiC_t^i : CPU utilization

QtiQ_t^i : Task queue length

EtiE_t^i : Energy level

DtiD_t^i : Average task deadline

L(t)iL_{\left( t \right)}^i: Network latency

This state is fed into a local reinforcement learning agent that selects the most appropriate scheduling action: execute the task locally, migrate to another edge node, or offload to the cloud. The agent is trained using reward functions that consider energy efficiency, task delay, successful completion, and load balancing.

The reinforcement learning agent updates its policy using a Deep Q-Network (DQN), defined by Equation (2): 2Q(Sti,ati)←Q(Sti,ati)+α[ rti+ϒmaxáQ(S(t+1)i,á)-Q(Sti,ati) ]Q\left( {S_t^i,a_t^i} \right) \leftarrow Q\left( {S_t^i,a_t^i} \right) + \alpha \left[ {r_t^i + \Upsilon ma{x_\'a }\;Q\left( {S_{(t + 1)}^i,\'a } \right) - Q\left( {S_t^i,a_t^i} \right)} \right]

Where

Q(Sti,ati)Q\left( {S_t^i,a_t^i} \right): Estimated Q-value for state-action pair

α: Learning rate

Y: Discount factor

rtir_t^i: Reward received for taking action atia_t^i

The concept of reward design plays a crucial role in shaping the learning behaviour of the FRL agent and so it is defined in detail. The reward at time t is computed using a multi-objective function that incorporates delay, energy consumption, SLA satisfaction and tardiness penalties. The overall reward is expressed as defined in Equation 3.3Rt=w1(−Dnorm)+w2(-Enorm)+w3SLAsuccess−w4Penaltytardiness,{R_t} = w1\left( { - {D_{norm}}} \right) + {w_2}\left( { - {E_{norm}}} \right) + {w_3}SL{A_{success}} - {w_4}{\rm{ Penalt}}{{\rm{y}}_{tardiness,}} where D_norm denotes the normalized task delay which presents the comparison of the actual delay with its allowable range. E_norm represents normalized energy usage which allows comparison of different energy consumption values on the same scale. SLA_success denotes whether a task completion has met its deadline or not represented using a binary or scaled indicator and Penalty_tardiess captures the degree to which a task completion exceeds its SLA-defined deadline. In order to maintain the numerical stability during the learning process, all performance variables are normalized via min–max scaling within each evaluation interval. This adaptive scaling ensures that the reward values remain within the range [0,1]. This normalization strategy prevents extreme fluctuations in reward magnitude, enabling smoother convergence of the FRL model.

If a task is completed before its deadline then it is assigned a positive value by the SLA component, otherwise it is assigned zero. As opposed to this, there is a proportional increase in the tardiness penalty with respect to the delay in deadline. This approach ensures that there is a strong penalty for time-sensitive applications thus making sure that agents are more responsive towards deadline-critical workloads. The selection of the weight parameters w₁, w₂, w₃, and w4determines the trade-offs between delay minimization, energy efficiency, SLA adherence and penalization severity for overdue tasks. A sensitivity analysis was conducted to check robustness of the system by varying each weight in the range of 0.1 to 0.9 while keeping the others constant. This analysis allowed the impact of different objective priorities to be evaluated systematically. The analysis revealed that the hybrid FRL–A-ABC model maintains stable performance across a broad range of weight configurations, demonstrating the adaptability of the proposed scheduling framework under diverse optimization preferences.

A fitness function that adapts with time is used by the ABC-optimizer to rearrange the task queue thus improving the granularity of local scheduling. This fitness function is based on some environmental factors such as energy drain rate and task deadline violations. The adaptive fitness function used by the ABC algorithm is expressed as follow in Equation (4): 4Fj=w1.lTj+w2.lEj+w3.Pj{F_j} = {w_l}.{l \over {{T_j}}} + {w_2}.{l \over {{E_j}}} + {w_3}.{P_j}

F_j: Fitness value of task scheduling strategy j

T_j: Average task completion time

E_j: Energy consumption

P_j: Priority score of the task queue

w₁, w₂, w₃: Weights determined by context (deadline urgency, load trend)

The employed swarm intelligence mimics the natural behavior of honeybees in foraging and dynamically balances exploration and exploitation of scheduling strategies. Each edge node executes this optimization independently ensuring low computational complexity and rapid convergence.

Federated learning serves as the backbone for knowledge sharing. Instead of transmitting raw data between nodes, model updates (i.e., the weights of the reinforcement learning agent) are periodically sent to a secure aggregator at the cloud layer. The cloud aggregator performs federated averaging to generate a global model as follow in Equation (5): 5θglobal=∑(i=1)Nnintotalθi{\theta _{global}} = \sum\nolimits_{(i = 1)}^N {{{{n_i}} \over {{n_{total}}}}} \;{\theta _i}

Where:

θ_i: Local model parameters from edge node iii

n_i: Number of tasks processed by node iii

ntotal=∑(i=1)Nni{n_{total}} = \sum\nolimits_{{\rm{(i}}\;{\rm{ = }}\;{\rm{1)}}}^N {ni} : Total tasks across all nodes

θ_global : Aggregated global model

This updated model is redistributed back to the edge agents. The periodicity and frequency of aggregation are adaptive and determined on the basis of local model divergence, communication bandwidth and task arrival rates.

A series of simulation experiments were conducted in order to validate the effectiveness of the proposed task scheduling technique. An enhanced CloudSim environment was used for conducting these experiments which was integrated with edge and fog computing capabilities [22]. Heterogeneous workloads and distributed computing resources were used so that the experimental setup would closely emulate the real-world edge cloud scenarios.

The simulation environment was configured as follows in Table 2.

To assess the performance of the proposed approach, it was compared with the following state-of-the-art scheduling algorithms: Ghanavati et al. [11]: Energy-aware Ant Colony Optimization (ACO) for fog computing environments. Ramezani et al. [14]: Reinforcement Learning-based load balancing across edge-fog-cloud layers. Shen et al. [17]: RL-based dynamic task scheduling in heterogeneous resource environments. The system’s performance was evaluated using the following quantitative metrics as listed in Table 3:

The effectiveness of the proposed FRL and Adaptive Swarm Intelligence-based scheduling framework was evaluated using the simulation environment described in the previous section. Several recent benchmark methods were taken from the literature [11, 16, 20] to do the performance evaluation of the proposed model. Results of the performance evaluation across multiple metrics demonstrated that the proposed hybrid model outperforms conventional reinforcement learning and bio-inspired strategies under varying conditions.

6.1

Task Completion Time (TCT)

The proposed framework demonstrated 13.4% lower average Task Completion Time (TCT) when compared to Ramezani et al. [16]. This significant improvement was made possible because of the federated learning mechanism used in the proposed model. This mechanism enabled the distributed edge agents to learn collaboratively and improve their optimal task scheduling strategies over time. The proposed model also showed better responsiveness to sudden bursts of dynamic task arrival as compared to the ant-based fog scheduling approach in Ghanavati et al. [11]. The proposed swarm intelligence component was capable of reordering the local task queues and hence reducing queuing delays and execution latency. Figure 3 presents an overview of task completion time comparison.

Figure 3.

Task completion time comparison

6.2

Energy Efficiency

The proposed model also incorporates adaptive intention filtering and decentralized decision making to optimize energy consumption across all edge and fog layers. Because of this, the proposed model was able to achieve 17% reduction in total energy consumption as compared to the baseline methods. The proposed model focuses on the computational constraints as well as the energy constraints of devices while scheduling. This results in more sustainable and eco-friendly decision making unlike Shen et al. [20]. Figure 4 provides the energy efficiency comparison of the proposed model with conventional methods.

Figure 4.

Energy efficiency comparison

6.3

Load Balancing

a significant improvement was shown by the proposed framework in terms of load distribution across all participating nodes. A decrease of 22% was shown in case of standard deviation of node workloads which indicates more uniform distribution of task assignments. As compared to the single-agent RL models described in [14, 17], the proposed collaborative framework enabled agents to anticipate traffic patterns and avoid hotspots by sharing gradient information which further lead to better congestion mitigation. Figure 5 provides an overview of the load balancing comparison.

Figure 5.

Load balancing comparison

6.4

Task Success Ratio (TSR)

Maintaining the task completion reliability is very important in case of highly dynamic and resource constrained environments. The proposed framework was able to achieve a task success ratio (TSR) of 94.3% which outperforms both techniques presented in [11, 16] that demonstrated a significant drop in the TSR under fluctuating mobility patterns or changing QoS demands. The continuous policy updates provided by FRL and the adaptive fitness driven reordering of tasks by ABC helped in achieving this robustness of the proposed framework. Figure 6 provides an overview of the task success ratio comparison.

Figure 6.

Task success ratio comparison

6.5

Scheduling Overhead

The FRL component might introduce some additional communication overhead but it is efficiently amortized through distributed agent training and lightweight coordination. The proposed framework demonstrated a 28% reduction in scheduling cost as compared to the actor-critic based models described in [12]. This was made possible because of the asynchronous updates and the low complexity nature of the ABC optimization at each node thus making the proposed framework highly suitable for real-time or near real-time IoT deployments. Figure 7 provides an overview of scheduling overhead comparison.

Figure 7.

Scheduling overhead comparison

6.6

Model Convergence

Lesser number of communication rounds were required by the proposed framework to converge on a global scheduling policy. On an average 50-60 training rounds are required for global convergence which is reportedly faster than the vehicular focussed models. This fast convergence was made possible in the proposed model because of the relatively lightweight task structures and distributed coordination thus ensuring rapid deployment and significant reduction in operational cost. Figure 8 provides an overview of model convergence comparison.

Figure 8.

Model convergence comparison

The authors in this paper have proposed a new hybrid task scheduling framework for edge cloud environments. The new framework presented in this paper is a combination of FRL and adaptive swarm intelligence algorithms in order to provide intelligent, energy-aware and context-aware task scheduling in heterogeneous cloud environments.

The methodology proposed in this paper allows edge agents to do privacy preserving collaboration which was not possible in traditional centralized or heuristic based methods. Because of this feature the agents can dynamically adapt to fluctuating workloads and resource constraints and also allows them to learn from distributed experiences. The experimental results demonstrated better performance as compared to recent reinforcement learning and bio inspired scheduling methods discussed in [11, 16, 20]. The parameters used for this comparison were task completion time, load balancing, energy consumption and task success ratio. The proposed model can be applied in real world domains such as industrial IoT, digital twin [23] smart cities [24], healthcare systems and edge AI applications [25] as its modular nature provides better scalability.

Stronger reinforcement learning baselines such as PPO, DDPG and SAC can be incorporated in the future work to make the experimental comparison more extensive. Multi-objective evolutionary algorithms like NSGA-II and MOEA-D can also be incorporated in the future work. These algorithms will allow a deeper performance evaluation of the proposed model under equal computational budgets. The simulation setup of the future experiments can also be extended by integrating richer workload models, detailed network and energy assumptions. More diverse edge-fog-cloud topology can also be used and additional statistical validation can be done using multiple independent runs and significance tests. Incorporation of all these enhancements will help in further strengthening the robustness and generalizability of the proposed framework.