A fusion deep Q-learning and particle swarm optimization algorithm for adaptive resource allocation in cloud computing circumstances Authors: Ahmed Hadi Ali AL-Jumaili, Mohammed E. Seno, Waleed Kareem Awad, Ravie Chandren Muniyandi, Mohammad Kamrul Hasan Published: 2026 ABSTRACT Effective resource allocation in cloud computing continues a critical challenge due to dynamic loads, stringent service-level expectations, and the need to balance execution time, energy, and cost. This study suggests a hybrid framework that integrates Deep Q-Learning (DQL) with Particle Swarm Optimization (PSO) to aid adaptive, multi-objective scheduling. DQL learns allocation strategies through interaction with the cloud environment, while PSO performs global search to refine action selection and accelerate convergence. Using Cloud Sim with real and synthetic workloads (Google Cluster, Planet Lab traces), the proposed method achieved a 35% reduction in average task execution time (from 245 s to 159 s) and a ~ 40% relative growth in resource utilization (from 60.1% to 84.6%), reduced SLA violations from 28 to 8, and lowered energy consumption to 6.3 kWh, outperforming standalone and hybrid models across 30 independent runs. Statistical tests (two-tailed t-test, α = 0.05) confirm significance. These results demonstrate that coupling reinforcement learning among swarm intelligence yields adaptive, high-quality decisions on behalf of real-time cloud resource scheduling. Introduction The high speed of the growth of cloud technology, it is now much easier for services to be scaled and accessed when needed by many industries1–3. As more mission-critical applications move to the cloud, the importance of having clever methods to allocate resources has grown very fast4. In the cloud, workloads can change a lot, SLAs can vary and deciding between performance, price and energy use is often complicated5. Classic approaches that depend on static and guesswork rules do not catch up with changes, so resources are not used wisely, processes take more time, and problems arise with meeting the expected levels of performance6. Current management approaches are not well suited to handling the unexpected and bursty pattern of workload that is frequently seen in cloud computing7. Most existing approaches do not respond well to different workloads and aim to minimize task time, make fuller use of resources and keep service level agreements in place8. Heuristics do not react to changes well, while standard machine learning models might not achieve the best result or stay reliable in every situation9. So, the cloud computing environment requires a resource allocation framework that is adaptive, intelligent and can respond in real time to its complex and unstable nature10,11. Recently, work has begun to blend reinforcement learning with metaheuristics for cloud scheduling, highlighting complementary strengths: RL adapts to grid-based policies, while metaheuristics accelerate global exploration and avoid local optima. For example, modified RL with VM consolidation and SA-PSO double Q learning have shown encouraging improvements, but still lack in scalability, joint policy hyperparameter optimization, and validation of real workloads. Based on these insights, this study designs a two-level DQL-PSO coupling (policy improvement + dynamic hyperparameter tuning) and evaluates it on Google Cluster and PlanetLab Trace to establish both performance and reproducibility. Three main contributions were made: creating a new hybrid model that depends on DQL and PSO for multi-objective scheduling, running tests on cloud workload data to confirm its usefulness and proving that it outperforms other recent algorithms in terms of completion time, better resource utilization and higher meetings of SLA metrics. They help improve cloud resource management by giving a useful, scalable way to apply theory to actual use in the real world. Two important approaches, Deep Q-Learning and Particle Swarm Optimization, are the main ideas in this research. Using DQL, deep reinforcement learning can help a system decide one action after another by experiencing the environment. Because it responds well to movements and uncertainties, it is rare that backprop is useful in tasks where what lies ahead isn’t known in advance. Alternatively, PSO focuses on exploring a big search area with stochastic simulations using social behaviors which helps it achieve results that are close to best solutions12. Here, we put forward a system where DQL makes most of the decisions for resource allocation by following feedback information and PSO globally optimizes these choices and fastens the convergence13. The combination of exploration and exploitation means the system can quickly change its actions in response to new work, while still performing well over time. Even though using learning and optimization for education works well in some domains, it hasn’t been fully explored in cloud resource management yet14,15. We propose a method that uses Deep Q-Learning (DQN) and Particle Swarm Optimization (PSO) for adaptive handling of resources in Cloud Computing (as seen in Fig. 1). Fig. 1DQN plus PSO fusion designed for adaptive resource allocation in cloud computing. DQN plus PSO fusion designed for adaptive resource allocation in cloud computing. It places Deep Q-Learning (DQN) and Particle Swarm Optimization (PSO) algorithms together, showing their use for adapting resources in a cloud setting. It outlines that in the cloud computing environment, DQN examines the system’s information and reward to figure out the right action for the agent. Using the policy it has learned, the agent chooses how to allocate resources which are then improved by PSO. PSO repositions particles to achieve the best way to spread resources and make sure everything is distributed well. In this diagram, DQN and PSO are shown to work together to achieve the best cloud resource management outcomes. Section II of this paper contains a review of studies that use classical as well as advanced resource allocation methods, reinforcement learning and swarm intelligence in cloud computing. Section III describes the starting points for DQL and PSO and explains how both are merged. Section IV explains the proposed hybrid algorithm, covering its structure, how it works and the official meanings of states, actions and rewards. In Section V, the authors describe the setting for the experiments, the data involved and how the results will be measured. The research results and a detailed discussion are included in Section VI and Section VII talks about what the method can do and its drawbacks moving forward. In Section VIII, the paper sums up its findings and underlines the value of hybrid intelligent systems in allocating resources on the cloud. Related work It is difficult to effectively use resources in cloud computing environments since workloads, service requirements and varieties of hardware and software are all always changing. The most common traditional scheduling schemes depend on straightforward techniques that do not adjust well when things change unexpectedly. Both FCFS and RR are common because of their ease in employment and because they are straightforward16. However, neither of these approaches is easy to scale or able to handle optimizing multiple criteria such as the time needed, the cost and energy at the same time. Because FCFS sorts of jobs based solely on the order they arrive, busy resources may be wasted, and people could wait longer. A large range of task durations can also reduce the efficiency of RR. Genetic Algorithms (GA) which form another kind of classical optimization, are being used to optimize resource scheduling by modeling nature’s selection and changes over generations17. GA-based methods are more efficient at searching the entire space and manage optimization involving several objectives much better than either FCFS or RR. However, they are not simple to run, and users must often make changes to the mutation rate and crossover probability to achieve good results. In addition, GAs struggle to react to sudden changes in workload, making them less practical in cloud environments that are always changing. Adaptive decision-making in uncertain environments now benefits from Reinforcement Learning (RL). RL helps agents discover the best ways to decide through many interactions with the environment which makes it ideal for tasks requiring multiple steps, including task scheduling, virtual machine location and load balancing18. By using DQL which adds deep neural networks to Q-learning algorithms, researchers have achieved good outcomes in modeling difficult state-action spaces and coping with rapid changes in the cloud environment19. DQL is used successfully across many areas of cloud computing. Researchers have applied DQL to better decide how and when to migrate VMs in data centers, save energy in cloud-edge networks and maintain excellent service in situations with changing consumer demand20–22. Without having to model the system, DQL adapts the way it acts as it receives input from the surroundings. Yet, straightforward RL models perform poorly in some areas: they can be unreliable when learning, are sensitive to the way rewards are given and need more time to improve in big environments. Failing these challenges early on can harm the performance of an information system. Robust methods for optimizing cloud resource management have been made popular by using Swarm Intelligence, a technique taken from biological group behavior. In SI algorithms, Particle Swarm Optimization (PSO) is very well known because it simulates particles in a search space to approach the best solutions step by step23,24. Many resource allocation issues are solved with PSO thanks to its quick rate of finding a solution and simple implementation. Cloud computing has used PSO to optimize the assigning of virtual machines, to schedule different tasks in multiple systems and to improve system reliability25,26. On average, PSO gets to convergence quicker and does so with less control than genetic algorithms. Yet, it sometimes converges too soon and can get stuck in certain ranges of complex data. It is also missing the ability to recall earlier experiences and so it depends only on existing particle coordinates for its progress. Table 1Comparative indication of resource allocation approaches in cloud computing.ApproachTechniqueAdvantagesLimitationsUse caseClassicalFCFS, Round RobinSimple, easy to implementPoor scalability, no optimizationSmall-scale static workloadsClassicalGenetic AlgorithmMulti-objective capability, good global searchHigh computation, slow convergenceStatic or semi-dynamic environmentsReinforcement LearningQ-Learning, DQN/DQLAdaptive, learns from experienceTraining instability, high varianceDynamic and unpredictable workloadsSwarm IntelligenceParticle Swarm OptimizationFast convergence, parallelizablePremature convergence, local optimaReal-time optimization problemsHybridDQL + PSOCombines adaptability and convergenceComplex design, tuning requiredLarge-scale, multi-objective resource allocationHybridGA + RLExploits global and local searchHigh computational overheadLong-term planning and scheduling Table 1Comparative indication of resource allocation approaches in cloud computing.ApproachTechniqueAdvantagesLimitationsUse caseClassicalFCFS, Round RobinSimple, easy to implementPoor scalability, no optimizationSmall-scale static workloadsClassicalGenetic AlgorithmMulti-objective capability, good global searchHigh computation, slow convergenceStatic or semi-dynamic environmentsReinforcement LearningQ-Learning, DQN/DQLAdaptive, learns from experienceTraining instability, high varianceDynamic and unpredictable workloadsSwarm IntelligenceParticle Swarm OptimizationFast convergence, parallelizablePremature convergence, local optimaReal-time optimization problemsHybridDQL + PSOCombines adaptability and convergenceComplex design, tuning requiredLarge-scale, multi-objective resource allocationHybridGA + RLExploits global and local searchHigh computational overheadLong-term planning and scheduling Comparative indication of resource allocation approaches in cloud computing. To solve the problems of single learning approaches, scientists have started to design systems that mix the best features of different ways of learning. Such models often make use of reinforcement learning along with swarm intelligence or evolutionary algorithms to control both exploration and exploitation and allow the agent to respond to changes in its environment27. Several experts have suggested using DQL together with PSO to boost both how fast policy convergence and how stable it remains in moving situations28. For multi-objective optimization of cloud resources, other studies have mixed fuzzy logic, genetic algorithms and neural networks29. Though they work well in tests, existing hybrid models have not generally been proven with real-world data, and they often struggle to scale up when used in many clouds at once30. Furthermore, there have been few attempts at adaptive reward shaping or dynamic parameter tuning in hybrid setups which means there is still room for improvement in this part of research. A serious challenge in design is deciding how much effort and resources to use against the benefits they offer, particularly with numerous layers of learning and optimization. For a clearer comparison of resource allocation strategies, Table 1 highlights the main characteristics, plus benefits and complications, of classical, reinforcement learning-based, swarm intelligence and hybrid ways. This comparative analysis highlights the need for intelligent hybridization that leverages the strengths of multiple methodologies while mitigating their individual weaknesses. The next section will delve into the theoretical foundations of DQL and PSO, explaining how these two techniques can be integrated into a unified framework for adaptive resource allocation in cloud computing. This paper introduces an algorithm that uses Deep Q-Learning and Particle Swarm Optimization to address the changing and difficult problem of resource allocation in the cloud. While other heuristic or standalone machine learning strategies cannot do both, our integration of DQL and PSO in the model allows the system to quickly adjust resources as needed and search the entire space to find the best solution. What makes the algorithm unique is the way reinforcement learning and swarm intelligence operate together, helping it manages exploration and exploitation more successfully than previous hybrids. We perform many experiments using data collected from real applications and with simulation tools like CloudSim to test the suggested approach. We assess the performance of our technique by examining it’s time for task execution, resource consumption, how often SLA rules are not followed and energy usage, as compared to individual algorithms (including DQL, PSO, Genetic Algorithm) and other hybrid methods. The testing process found that using this method can save 35% on the duration of the process and improve use of resources by 40%. Theoretical foundations Unlike standard Q-learning, where Q-values are directly estimated, Deep Q-Learning (DQL) depends on deep neural networks to approximate the function which makes it capable of handling many state variables in cloud computing situations31. In DQL, an agent interrelates with the conditions through a sequence of interpretation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}\in\:S$$\end{document}, actions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}\in\:A$$\end{document}, and rewards \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}\in\:\mathbb{R}$$\end{document}. The aim is to increase the expected cumulative reduced reward:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{t}=\sum\:_{k-0}^{{\infty\:}}\:{\gamma\:}^{k}{r}_{t+k+1}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:\in\:\left[\mathrm{0,1}\right]$$\end{document} is the reduction factor that determines how much the agent values upcoming rewards. The Q-value function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q\left({s}_{t},{a}_{t};\theta\:\right)$$\end{document} estimates the anticipated return of taking action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}$$\end{document} in state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document} using a deep neural network parameterized by weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}. The learning procedure involves decreasing the following loss event:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L\left(\theta\:\right)=\mathbb{E}\left[{\left({y}_{t}-Q\left({s}_{t},{a}_{t};\theta\:\right)\right)}^{2}\right]$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}={r}_{t}+\gamma\:\underset{{a}^{{\prime\:}}}{\mathrm{m}\mathrm{a}\mathrm{x}}\:Q\left({s}_{t+1},{a}^{{\prime\:}};{\theta\:}^{-}\right)$$\end{document}is the target Q-value computed using a separate goal network with boundaries \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}, which is updated continually to stabilize exercising. Metrics such as CPU usage, memory state, current job waiting time and SLA rules may be part of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document} element in cloud resource allocation context. Actions like \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}$$\end{document} include decisions to assign a VM or container to a particular host. The reward function ensures multiple optimization goals are balanced properly which are usually set as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}={w}_{1}\cdot\:\frac{1}{{T}_{\mathrm{exec\:}}}+{w}_{2}\cdot\:{U}_{\mathrm{res\:}}-{w}_{3}\cdot\:{V}_{SLA}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\mathrm{exec\:}}$$\end{document} is the total time spent executing, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{\mathrm{res\:}}$$\end{document} measures the average use of resources, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{SLA}$$\end{document} counts SLA regulations that are not met and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1},{w}_{2},{w}_{3}$$\end{document} are used to rank these metrics differently32. Particle Swarm Optimization (PSO) uses a behavior-like method inspired by animals moving together to solve problems33. Every particle shows a possible solution and moves to a new position guided by what it has learned, and the best spots found by itself and by other particles in the swarm. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}=\left({x}_{i1},{x}_{i2},\dots\:,{x}_{id}\right)$$\end{document} be the location vector of particle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} in a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d$$\end{document}-dimensional space, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}=\left({v}_{i1},{v}_{i2},\dots\:,{v}_{id}\right)$$\end{document} be its velocity vector. Each particle revises its location and velocity allowing to the subsequent rules:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{v}_{id}^{(t+1)}=\omega\:{v}_{id}^{\left(t\right)}+{c}_{1}{r}_{1}\left({p}_{id}-{x}_{id}^{\left(t\right)}\right)+{c}_{2}{r}_{2}\left({g}_{d}-{x}_{id}^{\left(t\right)}\right)\\\:{x}_{id}^{(t+1)}={x}_{id}^{\left(t\right)}+{v}_{id}^{(t+1)}\end{array}$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} is the inertia weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{2}$$\end{document} are cognitive and social increase in speed constants, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{2}$$\end{document} are random integers uniformly distributed in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[\mathrm{0,1}\right]$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{id}$$\end{document} is the best location of particle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{d}$$\end{document} is the best global locate among all particles. To ensure convergence while preserving exploration, the inertia weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} is often linearly diminished over repetitions34:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:\left(t\right)={\omega\:}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\left(\frac{{\omega\:}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\omega\:}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{T}_{\mathrm{m}\mathrm{a}\mathrm{x}}}\right)t,$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} is the limit number of repetitions. Because PSO finds good answers quickly and can explore a wide range of solutions, it is widely used in cloud computing for VM placement, scheduling tasks and load balancing35. Different previous PSO-guided DQN methods, which limit PSO to action enhancement, and SA-PSO-Q learning variants that add stochastic temperature control, the proposed framework introduces a two-level coupling: (i) policy-level action refinement, where PSO searches for top-k DQL-ranked actions to improve immediate rewards; and (ii) online hyperparameter tuning, where PSO dynamically adjusts key DQL parameters (learning rate α, relaxation γ, exploration ε) to stabilize training and accelerate convergence. In addition, we (iii) shape the reward to balance performance time, utilization, SLA and energy, and (iv) validate on real-world traces (Google Cluster, PlanetLab) in addition to artificial weights. It jointly addresses edge optimization (DQL) and global convergence (PSO), achieving faster convergence, lower variance, and better generalization under intense workloads. Our study develops a twofold framework in which DQL leads the process of making decisions, developing policies and adapting to the environment, while PSO adds value by tuning actions and hyperparameters36. Specifically, the PSO component operates at two levels: Case Space Optimization: For a given state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}, instead of selecting the longing action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{a}{\mathrm{m}\mathrm{a}\mathrm{x}}\:Q\left({s}_{t},a;\theta\:\right)$$\end{document}, PSO explores the area of the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions predicted by DQL to find possibly better results.Parameter Regulation: PSO is exercised to dynamically improve key hyperparameters of the DQL model, such as the learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document}, concession factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document}, and search rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon$$\end{document}, enhancing merging and strength. Case Space Optimization: For a given state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}, instead of selecting the longing action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{a}{\mathrm{m}\mathrm{a}\mathrm{x}}\:Q\left({s}_{t},a;\theta\:\right)$$\end{document}, PSO explores the area of the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions predicted by DQL to find possibly better results. Case Space Optimization: For a given state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}, instead of selecting the longing action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{a}{\mathrm{m}\mathrm{a}\mathrm{x}}\:Q\left({s}_{t},a;\theta\:\right)$$\end{document}, PSO explores the area of the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions predicted by DQL to find possibly better results. Parameter Regulation: PSO is exercised to dynamically improve key hyperparameters of the DQL model, such as the learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document}, concession factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document}, and search rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon$$\end{document}, enhancing merging and strength. Parameter Regulation: PSO is exercised to dynamically improve key hyperparameters of the DQL model, such as the learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document}, concession factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document}, and search rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon$$\end{document}, enhancing merging and strength. The overall manipulation loop movements are as follows: At each time step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document}, the DQL agent keeps state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document} and creates a ranked list of candidate actions. PSO values these applicants in parallel and selects the one that maximizes the assumed reward. The chosen action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}$$\end{document} is executed, and the resulting move \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({s}_{t},{a}_{t},{r}_{t},{s}_{t+1}\right)$$\end{document} is stored in the repeat buffer. Regularly, the PSO module adjusts DQL’s hyperparameters based on recent execution metrics. This hybrid planning ensures that the system grants from both the long-term flexibility of DQL and the short-term optimization power of PSO. It helps eschew local optima, increases meeting speed, and enhances robustness against dynamic workload variations, making it predominantly well-suited for real-time source allocation in cloud locations. Algorithm 1Mix DQL-PSO Resource Allocation. Mix DQL-PSO Resource Allocation. Step 1: Create Candidate Actions With DQL: The DQL agent computes the Q-values for all possible actions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a\in\:A$$\end{document} in the existent state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}. From these Q -values, the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions are chosen as contender movements for advance discovery.Step 2: Optimize Acts With PSO: A swarm of elements is modified in the getting space, where each particle characterizes a potential solution (action). For each particle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, its velocity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} and stance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} are revised relating to the PSO equations: Step 1: Create Candidate Actions With DQL: The DQL agent computes the Q-values for all possible actions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a\in\:A$$\end{document} in the existent state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}. From these Q -values, the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions are chosen as contender movements for advance discovery. Step 1: Create Candidate Actions With DQL: The DQL agent computes the Q-values for all possible actions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a\in\:A$$\end{document} in the existent state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t}$$\end{document}. From these Q -values, the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document} actions are chosen as contender movements for advance discovery. Step 2: Optimize Acts With PSO: A swarm of elements is modified in the getting space, where each particle characterizes a potential solution (action). For each particle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, its velocity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} and stance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} are revised relating to the PSO equations: Step 2: Optimize Acts With PSO: A swarm of elements is modified in the getting space, where each particle characterizes a potential solution (action). For each particle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, its velocity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} and stance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} are revised relating to the PSO equations: Velocity update:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{id}^{(t+1)}=\omega\:{v}_{id}^{\left(t\right)}+{c}_{1}{r}_{1}\left({p}_{id}-{x}_{id}^{\left(t\right)}\right)+{c}_{2}{r}_{2}\left({g}_{d}-{x}_{id}^{\left(t\right)}\right)$$\end{document} Rank update:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{id}^{(t+1)}={x}_{id}^{\left(t\right)}+{v}_{id}^{(t+1)}$$\end{document} The strength of each particle is weighed based on the reward function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r\left({x}_{i}\right)$$\end{document}. After realizing the PSO repetitions, the best activity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}^{\mathrm{*}}$$\end{document} is chosen from the optimized nominations. 3.Step 3: Perform Act and View Next State: The select action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}^{\mathrm{*}}$$\end{document} is completed in the natural environment, subsequent in a new state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document} and an immediate reward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}$$\end{document}.4.Step 4: Store Change in Replay Buffer: The transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({s}_{t},{a}_{t}^{\mathrm{*}},{r}_{t},{s}_{t+1}\right)$$\end{document} is stored in the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. This buffer is used to sample mini batches for training the DQL network.5.Step 5: Train DQL Network: A mini batch of changes is sampled from the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. Target Q-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}$$\end{document} are computed using the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}. The DQL network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} are updated to minimize the loss function: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L\left(\theta\:\right)=\mathbb{E}\left[{\left({y}_{t}-Q\left({s}_{t},{a}_{t};\theta\:\right)\right)}^{2}\right]$$\end{document} 3.Step 3: Perform Act and View Next State: The select action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}^{\mathrm{*}}$$\end{document} is completed in the natural environment, subsequent in a new state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document} and an immediate reward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}$$\end{document}. Step 3: Perform Act and View Next State: The select action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t}^{\mathrm{*}}$$\end{document} is completed in the natural environment, subsequent in a new state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document} and an immediate reward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}$$\end{document}. 4.Step 4: Store Change in Replay Buffer: The transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({s}_{t},{a}_{t}^{\mathrm{*}},{r}_{t},{s}_{t+1}\right)$$\end{document} is stored in the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. This buffer is used to sample mini batches for training the DQL network. Step 4: Store Change in Replay Buffer: The transition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({s}_{t},{a}_{t}^{\mathrm{*}},{r}_{t},{s}_{t+1}\right)$$\end{document} is stored in the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. This buffer is used to sample mini batches for training the DQL network. 5.Step 5: Train DQL Network: A mini batch of changes is sampled from the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. Target Q-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}$$\end{document} are computed using the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}. The DQL network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} are updated to minimize the loss function: Step 5: Train DQL Network: A mini batch of changes is sampled from the replay buffer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B$$\end{document}. Target Q-values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}$$\end{document} are computed using the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}. The DQL network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} are updated to minimize the loss function: 6.Step 6: Periodically Revise Target Network: To stabilize exercise, the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}are periodically informed to match the primary network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}.7.Step 7: Update PSO Hyperparameters. 6.Step 6: Periodically Revise Target Network: To stabilize exercise, the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}are periodically informed to match the primary network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}. Step 6: Periodically Revise Target Network: To stabilize exercise, the target network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{-}$$\end{document}are periodically informed to match the primary network parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}. 7.Step 7: Update PSO Hyperparameters. Step 7: Update PSO Hyperparameters. The inertia weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} is dynamically adjusted using:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:\left(t\right)={\omega\:}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\left(\frac{{\omega\:}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\omega\:}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{T}_{\mathrm{m}\mathrm{a}\mathrm{x}}}\right)t$$\end{document} This ensures a balance between exploration and exploitation during PSO iterations. 8.Step 8: Update Current State: The current state is informed to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document}, and the process continues until the end of the episode. 8.Step 8: Update Current State: The current state is informed to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document}, and the process continues until the end of the episode. Step 8: Update Current State: The current state is informed to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{t+1}$$\end{document}, and the process continues until the end of the episode. To estimate the performance of the proposed hybrid DQL-PSO algorithm for adaptive resource allocation in cloud computing surroundings, a comprehensive set of simulation experiments was conducted using standard benchmark tools such as CloudSim. The model setup incorporates accurate cloud infrastructure configurations, dynamic workload patterns, and key algorithmic parameters for both the DQL and PSO components. Table 2 reviews the main parameters used in the experiments, including environment settings, knowledge parameters, swarm optimization settings, and performance estimate metrics. Table 2Simulation parameters and configurations.Parameter categoryParameter nameValue / rangeDescriptionEnvironment SettingsNumber of VMs50Total number of virtual machines available for task allocationNumber of Hosts20Physical servers hosting VMsCPU Capacity per Host16,000 MIPSProcessing capacity of each hostRAM per Host32 GBMemory size of each hostBandwidth1 GbpsNetwork bandwidth between hostsTask Arrival PatternPoissonSimulates random and dynamic workloadTask Length1000–5000 MI (Million Instructions)Range of computational requirements per taskDQL ParametersLearning Rateα0.001Controls the step size during neural network updatesDiscount Factorγ0.95Determines how much the agent values future rewardsExploration RateϵStarts at 1.0, decays to 0.1Probability of choosing random action vs greedy actionExperience Replay Buffer Size10,000Maximum number of transitions storedMini-batch Size64Number of samples used per DQL updateTarget Network Update FrequencyCEvery 100 stepsInterval for syncing target and main networksNeural Network Layers[Input: state size, Hidden: 128, Output: actions]Structure of the deep Q-networkPSO ParametersPopulation Size30Number of particles in the swarmCognitive Coefficientc1​1.5Personal best influenceSocial Coefficientc2​1.5Global best influenceInertia WeightωLinearly decreases from 0.9 to 0.4Balances exploration and exploitationMax IterationsTmax​100Maximum number of iterations per PSO runSearch Space DimensionsEqual to action space sizeEach particle represents an action vectorPerformance MetricsTask Execution TimeMeasured in secondsTime taken to complete all tasksResource Utilization RateTotal ResourcesUsed Resources​Percentage of utilized resourcesSLA Violation RateCount of SLA violations / total tasksMeasures adherence to service-level agreementsEnergy ConsumptionEstimated based on CPU usageEnergy consumed by hosts during task executionConvergence SpeedTraining episodes until stable policyIndicates how fast the model learns optimal decisions Table 2Simulation parameters and configurations.Parameter categoryParameter nameValue / rangeDescriptionEnvironment SettingsNumber of VMs50Total number of virtual machines available for task allocationNumber of Hosts20Physical servers hosting VMsCPU Capacity per Host16,000 MIPSProcessing capacity of each hostRAM per Host32 GBMemory size of each hostBandwidth1 GbpsNetwork bandwidth between hostsTask Arrival PatternPoissonSimulates random and dynamic workloadTask Length1000–5000 MI (Million Instructions)Range of computational requirements per taskDQL ParametersLearning Rateα0.001Controls the step size during neural network updatesDiscount Factorγ0.95Determines how much the agent values future rewardsExploration RateϵStarts at 1.0, decays to 0.1Probability of choosing random action vs greedy actionExperience Replay Buffer Size10,000Maximum number of transitions storedMini-batch Size64Number of samples used per DQL updateTarget Network Update FrequencyCEvery 100 stepsInterval for syncing target and main networksNeural Network Layers[Input: state size, Hidden: 128, Output: actions]Structure of the deep Q-networkPSO ParametersPopulation Size30Number of particles in the swarmCognitive Coefficientc1​1.5Personal best influenceSocial Coefficientc2​1.5Global best influenceInertia WeightωLinearly decreases from 0.9 to 0.4Balances exploration and exploitationMax IterationsTmax​100Maximum number of iterations per PSO runSearch Space DimensionsEqual to action space sizeEach particle represents an action vectorPerformance MetricsTask Execution TimeMeasured in secondsTime taken to complete all tasksResource Utilization RateTotal ResourcesUsed Resources​Percentage of utilized resourcesSLA Violation RateCount of SLA violations / total tasksMeasures adherence to service-level agreementsEnergy ConsumptionEstimated based on CPU usageEnergy consumed by hosts during task executionConvergence SpeedTraining episodes until stable policyIndicates how fast the model learns optimal decisions Simulation parameters and configurations. Proposed hybrid algorithm This proposed work combines DQL and PSO into one algorithm to adapt cloud computing resources more effectively. In our PSO, the system architecture incorporates four main buildings blocks: the environment interface, DQL agent, PSO optimizer and resource manager. The environment interface is constantly connected to the cloud infrastructure, capturing in real time CPU usage, memory levels, active tasks in the queue and SLA rules. The information is turned into the current state vector and given to the DQL agent for action. After analyzing suitable actions with learned rules, the DQL agent assembles a list of the strongest possibilities. The PSO optimizer looks over the candidates and performs more searches to improve the action selection process, using both innovative and effective methods. After that, the resource manager applies the chosen action and updates all needed information. A sequential and interactive loop joins reinforcement learning with swarm optimization in the workflow of the model. At first, the cloud system takes in new work and holds it in a queue to handle later. At each step, the system records its state in the environment interface and the DQL agent uses this information to determine the Q-values for every action. According to these values, the k best candidate actions are entered into the PSO optimizer. At the same time, the PSO module places a group of particles representing potential options and it gradually adjusts them to see which one performs the best. If the optimal action is chosen, the resource manager carries it out, then the replay buffer remembers the previous state, action taken, reward gotten and next state. Frequently, the DQL agent retrieves minibatches from the buffer for updating its policy network. Its target network is then updated periodically to maintain a stable learning process. Besides, the PSO optimizer changes its parameters, most importantly the inertia weight to maintain and even search across the entire problem space. For learning and optimization to work well, the hybrid model focuses on three important elements: state, action and reward. Within S, there are real-time metrics tracking CPU capacity, available memory, virtual machines in operation, the queue for current tasks and whether SLA conditions are complied with. These tools give the DQL agent a complete understanding of the cloud environment. Action space A involves making discrete choices about resources using tasks such as allocating to a VM or host, relocating VMs or modifying resource parameters. All actions contribute to how the system evolves and provide you with instant feedback. The objective of the reward function is to achieve the best possible time to run, the most balanced use of resources and strict adherence to service level agreements all at once. It is understood as:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}={w}_{1}\cdot\:\frac{1}{{T}_{\mathrm{exec\:}}}+{w}_{2}\cdot\:{U}_{\mathrm{res\:}}-{w}_{3}\cdot\:{V}_{SLA}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\mathrm{exec\:}}$$\end{document} denotes total execution time, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{\mathrm{res\:}}$$\end{document} denotes average source utilization, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{SLA}$$\end{document} indicates the number of SLA damages, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1},{w}_{2},{w}_{3}$$\end{document} are tunable weights exhibiting the relative weight of each objective. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} - Discount factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(0<\gamma\:<1)$$\end{document}, weighting future rewards. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{s},{\mathrm{s}}^{{\prime\:}}$$\end{document} - Current and next environment states (CPU, RAM, queue length, SLA flags). a - Discrete action (VM/host assignment, VM migration, parameter adjustment). r - Immediate reward computed from the multi-objective schedule outcome. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Q}(\mathrm{s},\mathrm{a};\theta\:)$$\end{document} - Action-value function approximated by a deep neural network with weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:,{\theta\:}^{-}$$\end{document}- Online and target network parameters (DQL). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} - Learning rate for gradient updates in DQL. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document} - Exploration probability for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}-greedy policy. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} - PSO inertia weight (linearly decays from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} _max to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} _min). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{c}}_{1},{\mathrm{c}}_{2}$$\end{document} - PSO cognitive and social coefficients. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{1},{r}_{2}$$\end{document} - Uniform random variables in [0,1] for PSO velocity update. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{p}}_{\mathrm{i}}$$\end{document} - Best position found by particle i (personal best). g - Best global position among all particles (global best). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{v}}_{\mathrm{i}},{\mathrm{x}}_{\mathrm{i}}$$\end{document} - Velocity and position vectors of particle i. k - Number of top actions from DQL used to seed PSO search. w_time, w_util, w_SLA, w_energy - Reward weights for time, utilization, SLA, and energy terms. This ensures that the agent is compensated for faster performance, greater application, and fewer SLA breaches. This algorithm divides its strategy into steps by alternately using reinforcement learning and swarm-based optimization. After every step, the DQL agent selects the best action it thinks the robot can take and the PSO optimizer improves this choice so the robot can learn and perform well. The replay buffer stores information about what happened in the environment and the algorithm updates the target network from time to time. Figure 2 shows the Technique of this study. demonstrates the back-and-forth exchange between the DQL agent and PSO optimizer in the total system. The first stage is watching over a state, then producing an action using DQL, optimizing based on PSO, running the best action and lastly, informing both learning and parameters. Because of its closed-loop structure, our solution can easily handle different and shifting patterns of workload and resource needs, making it highly valuable in cloud environments. Both data obtained from real systems and benchmark data were used to examine the proposed hybrid Deep Q-Learning and Particle Swarm Optimization (DQL-PSO) method. They were chosen so that the testing could cover many regular situations in cloud computing such as changes and unusual loads. Data availability Google Cluster Data was selected as the main workload source for our analysis. The full logs in this dataset were gathered from Google’s batch processing systems and include information on when tasks are sent, how long they take, their resource demands and if they finished properly. It represents the overall activities in data centers accurately and is an excellent reference for comparing scheduling algorithms under difficult and changing workloads. People can find the dataset here: https://github.com/google/cluster-data . Along with Google Cluster Data, the research used PlanetLab Virtual Machine Traces dataset. The dataset tracks how much CPU and memory each virtual machine uses when running on PlanetLab nodes worldwide. Because it represents real-world changes in resource demand and moments of inactivity, it is very helpful for studying how effectively the algorithm responds to changes in system load and differences in machine capabilities. Traces can be found at this link: Fig. 2Proposed hybrid LSTM–fuzzy-GWO–Q-learning framework for intelligent WBAN operation. Proposed hybrid LSTM–fuzzy-GWO–Q-learning framework for intelligent WBAN operation. https://www.mlr-lab.org/teaching-material/cloudsim/planetlab_dataset.zip. A series of synthetic workloads were also created using statistical properties to check the algorithm’s strength in extreme conditions. The model used the Poisson distribution to model task arrivals as unpredictable and occasional, and it used an exponential distribution to simulate various task lengths. Experiments with these synthesized datasets permitted controlled study of the technique under various scenarios, giving information on its response and adaptability. Experimental setup All experiments to check the hybrid DQL-PSO algorithm’s performance used the CloudSim simulation framework to simulate virtual computing resources. The simulated environment uses several data centers and hosts in each center are not the same in CPU and memory. The intelligent agent schedules tasks onto virtual machines that are set up on the fly as they are needed. The basic features and designs used in the simulation environment are detailed in Table 3. All experiments were performed on a workstation with an Intel Core i9-13900 K (24 cores, 3.5 GHz), 64 GB RAM, NVIDIA RTX 4090 (24 GB), Windows 11 and CloudSim 4.0. Each configuration was performed 30 independent runs; We report mean ± standard deviation. Statistical significance for pairwise comparisons with baseline was assessed using a two-tailed Student’s t-test (α = 0.05), and 95% confidence intervals are shown where appropriate. The simulation horizon per run was 6 h of cloud operation, with 500–2000 tasks generated via Poisson arrival; The length of the tasks follows an exponential distribution in accordance with previous work. Table 3Cloud simulation environment configuration.ComponentConfiguration DetailsSimulation ToolCloudSim 4.0Number of Data Centers2Hosts per Data Center10CPU Capacity per Host16,000 MIPSRAM per Host32 GBVM Instances per Host5VM CPU Capacity2,000–4,000 MIPS (randomly assigned)VM RAM4–8 GB (randomly assigned)Network Bandwidth1 Gbps between hosts Table 3Cloud simulation environment configuration.ComponentConfiguration DetailsSimulation ToolCloudSim 4.0Number of Data Centers2Hosts per Data Center10CPU Capacity per Host16,000 MIPSRAM per Host32 GBVM Instances per Host5VM CPU Capacity2,000–4,000 MIPS (randomly assigned)VM RAM4–8 GB (randomly assigned)Network Bandwidth1 Gbps between hosts Cloud simulation environment configuration. Both real-world and synthetic testing were performed by using two distinct categories of scaled workload datasets for the model. 1. Actual workloads gathered from the real world such as: Google Cluster Data includes full records of tasks running and resources used in Google’s internal clusters. CPU and memory activities of virtual machines are captured by PlanetLab VM Traces. 2. For simulation, I made workloads consisting of tasks launched in Poisson order and whose lengths followed exponential functions to create heavy and bursty traffic scenarios during testing. Thanks to this data, we could test how well the algorithm adapts, responds and generalizes to many different workload scenarios. To validate the effectiveness of the proposed hybrid approach, it was compared against several established resource allocation strategies, including standalone reinforcement learning, swarm intelligence, heuristic-based, and hybrid algorithms. All baseline models were implemented and evaluated under identical experimental settings. Both real-world and generated workloads were chosen for evaluation to show how the proposed DQL-PSO approach handles challenging and realistic cloud problems (Table 4 details all the data and baseline methods used). The conversion was assessed by looking at execution efficiency, the resources used and if service-level standards were met (Table 5). All algorithms were built using proper and adjusted parameters obtained from both experimental observations and previous studies (Table 6). The proposed algorithm was assessed against some popular cloud performance evaluation metrics, so the public could see how it performs, achieves high quality and can be adapted to changing needs. For both the DQL and PSO modules, the hyperparameters were optimized to guarantee the system converged and performed well. We set these settings based on early experiments and the process of cross-validation. This well-defined experiment lets us fully assess how the hybrid DQL-PSO algorithm compares with other algorithms and operates in several different cloud computing conditions. Results and analysis The proposed hybrid DQL-PSO algorithm was evaluated against several baseline methods across multiple analyses of its performance metrics to determine the success of its adaptive resource management. The results in Table 7 demonstrate that the hybrid model came out ahead of both component and hybrid algorithms for executing tasks, using available resources, violating SLAs and reducing energy usage. It became evident that choosing hybrid production yielded up to 35% fewer minutes for task execution than single-solution DQL, while also saving 40% on resources versus the old PSO method. Thankfully, these gains result from the unification of reinforcement learning for changing policies and PSO for searching for the best possible action selections. To validate the observed performance improvements, we conducted statistical hypothesis test results come from 30 repeated runs of each algorithm, all under the same workload conditions. Under the null hypothesis, the performance of the proposed method was not significantly different than the existing baselines. Since the p-values fell below α = 0.05, we can say that the increases in execution time, compliance to set SLA and effective use of resources confirmed by the hybrid DQL-PSO algorithm were significant results. As an example, compared with standalone DQL, our proposed method was found to be superior with p < 0.001. We went on to check how sensitive the algorithm was to changes in the intensity of workloads, number of VMs and capacity of hosts. Increasing the number of VMs to 60 as seen in Table 8., improves all performance measures and the DQL-PSO method showed the steadiest results. During stressful workloads (500 simultaneous tasks), the hybrid system kept up a good SLA at 12%, in comparison to DQL’s rate of 34%. This highlights that the algorithm can respond well to new changes in its environment, without having to be retrained. Table 4Workload datasets overview.Dataset typeNameDescriptionUse caseReal-worldGoogle Cluster DataLogs of job execution and resource usage from Google clustersRealistic dynamic workload modelingReal-worldPlanetLab VM TracesCPU and memory utilization of VMs across PlanetLab nodesResource behavior analysisSyntheticPoisson Task ArrivalTasks arrive randomly at an average rateHigh concurrency testingSyntheticExponential Task LengthTask durations follow exponential distributionBursty workload simulation Table 4Workload datasets overview.Dataset typeNameDescriptionUse caseReal-worldGoogle Cluster DataLogs of job execution and resource usage from Google clustersRealistic dynamic workload modelingReal-worldPlanetLab VM TracesCPU and memory utilization of VMs across PlanetLab nodesResource behavior analysisSyntheticPoisson Task ArrivalTasks arrive randomly at an average rateHigh concurrency testingSyntheticExponential Task LengthTask durations follow exponential distributionBursty workload simulation Workload datasets overview. Table 5Baseline methods for comparison.CategoryAlgorithmDescriptionReinforcement LearningDeep Q-Learning (DQL)Pure DQN-based policy without optimizationSwarm IntelligenceParticle Swarm Optimization (PSO)PSO-only scheduler with no RL componentEvolutionary AlgorithmGenetic Algorithm (GA)Population-based search method for task schedulingHeuristic MethodRound Robin (RR)Cyclic assignment of tasks to VMsHeuristic MethodFirst-Come-First-Served (FCFS)Simple queue-based schedulingHybrid ApproachDQL-GADQL with GA-based parameter tuningHybrid ApproachPSO-GAPSO combined with genetic selectionHybrid ApproachPSO-DQNPSO-guided action selection with DQN Table 5Baseline methods for comparison.CategoryAlgorithmDescriptionReinforcement LearningDeep Q-Learning (DQL)Pure DQN-based policy without optimizationSwarm IntelligenceParticle Swarm Optimization (PSO)PSO-only scheduler with no RL componentEvolutionary AlgorithmGenetic Algorithm (GA)Population-based search method for task schedulingHeuristic MethodRound Robin (RR)Cyclic assignment of tasks to VMsHeuristic MethodFirst-Come-First-Served (FCFS)Simple queue-based schedulingHybrid ApproachDQL-GADQL with GA-based parameter tuningHybrid ApproachPSO-GAPSO combined with genetic selectionHybrid ApproachPSO-DQNPSO-guided action selection with DQN Baseline methods for comparison. Table 6Performance evaluation metrics.Metric categoryMetric nameDefinition / measurementEfficiencyTask Execution TimeTotal time taken to complete all tasks (in seconds)Quality of ServiceResource Utilization RateTotal Available ResourcesUsed Resources​Quality of ServiceSLA Violation CountNumber of tasks that exceed their deadline or fail to meet SLA requirementsEnergy EfficiencyEnergy ConsumptionEstimated using host power consumption during active and idle statesLearning BehaviorConvergence SpeedNumber of training episodes until stable policy is learnedLearning BehaviorStabilityVariance in performance across multiple runs and changing workloads Table 6Performance evaluation metrics.Metric categoryMetric nameDefinition / measurementEfficiencyTask Execution TimeTotal time taken to complete all tasks (in seconds)Quality of ServiceResource Utilization RateTotal Available ResourcesUsed Resources​Quality of ServiceSLA Violation CountNumber of tasks that exceed their deadline or fail to meet SLA requirementsEnergy EfficiencyEnergy ConsumptionEstimated using host power consumption during active and idle statesLearning BehaviorConvergence SpeedNumber of training episodes until stable policy is learnedLearning BehaviorStabilityVariance in performance across multiple runs and changing workloads Performance evaluation metrics. Table 7Comparative performance across metrics.AlgorithmAvg. execution time (s)Resource utilization (%)SLA violationsEnergy consumption (kWh)Round Robin28762.4429.2FCFS30160.1459.5DQL24571.8288.1PSO23973.5257.9GA26068.3318.4Hybrid DQL-GA22875.2207.5Hybrid PSO-GA22277.4187.2Proposed DQL-PSO15984.686.3 Table 7Comparative performance across metrics.AlgorithmAvg. execution time (s)Resource utilization (%)SLA violationsEnergy consumption (kWh)Round Robin28762.4429.2FCFS30160.1459.5DQL24571.8288.1PSO23973.5257.9GA26068.3318.4Hybrid DQL-GA22875.2207.5Hybrid PSO-GA22277.4187.2Proposed DQL-PSO15984.686.3 Comparative performance across metrics. One of the key advantages of the hybrid DQL-PSO model lies in its faster convergence speed compared to standalone reinforcement learning methods. As depicted in Figure 2, the hybrid algorithm reached a stable policy within approximately 300 training episodes, while DQL alone required over 600 episodes to converge. Moreover, the integration of PSO significantly reduced variance in early learning stages, leading to more consistent updates and fewer oscillations during training. This indicates that the PSO component enhances the exploration phase, allowing the agent to discover better policies more efficiently. To demonstrate the real-world applicability and robustness of the proposed algorithm, we conducted a scenario-based analysis involving bursty workload patterns and dynamic VM provisioning. In one representative case, a sudden surge in incoming tasks caused CPU utilization to exceed 90% on several hosts. While heuristic methods like RR and FCFS failed to respond effectively, leading to increased SLA violations, the hybrid DQL-PSO model dynamically redistributed tasks to underutilized hosts, maintaining an average CPU usage of around 78% and minimizing downtime. This case highlights the algorithm’s ability to adapt to real-time changes and maintain service quality even under unexpected load fluctuations. This comprehensive evaluation confirms that the proposed hybrid DQL-PSO algorithm not only outperforms existing methods in standard benchmarks but also exhibits strong statistical significance, parameter robustness, fast convergence, and practical adaptability in real cloud computing scenarios. Training the proposed hybrid needs ≈ 7.3 min per 100 episodes, compared to 5.1 min for standalone DQL and 3.9 min for PSO. This ~ 30% training overhead is offset by faster convergence (≈ 300 vs. > 600 episodes for DQL alone) and higher final performance. Importantly, the inference delay per decision (by selecting the allocation action from the policy learned with PSO bounding) remained < 35 ms, thus meeting the real-time scheduling constraints in the simulated settings. Figure 3 illustrates the sensitivity of various scheduling models DQL, PSO, GA, and the proposed DQL-PSO to changes in system parameters such as workload intensity, the number of virtual machines (VMs), and host capacity. The top panel of the chart compares the execution time, while the bottom panel shows the corresponding SLA violation rates for each configuration. As evident from the chart, the DQL-PSO model consistently outperforms traditional and hybrid models, demonstrating the lowest execution time and SLA violation rates across all tested parameters. Particularly under high workload intensity (500 concurrent tasks), DQL-PSO achieves a remarkable reduction in SLA violation rate from 34% (DQL) to just 12%. Similarly, increasing VMs and host capacity yields further gain, underlining the scalability and robustness of the DQL-PSO approach. To provide a deeper understanding of the dynamics and stability of the hybrid model, two complementary analyses were conducted. Figure 4 shows the convergence behavior of the proposed DQL-PSO framework compared to the standalone DQL and PSO models, highlighting its faster convergence and smoother reward growth. Meanwhile, Fig. 5 presents the parameter sensitivity analysis, which shows how variations in the learning rate (α) and the inertia weights (ω) affect the average execution time. The results confirm that the hybrid model achieves both stable and efficient learning over a wide hyperparameter range and ensures reproducibility and adaptability to different cloud working conditions. Fig. 3Effect of system parameters on performance time and SLA violation rate across models. Effect of system parameters on performance time and SLA violation rate across models. Fig. 4Convergence curve of DQL–PSO vs. baselines. Convergence curve of DQL–PSO vs. baselines. Fig. 5Parameter sensitivity analysis of DQL–PSO (Learning Rate α vs. Inertia Weight ω). Parameter sensitivity analysis of DQL–PSO (Learning Rate α vs. Inertia Weight ω). Table 8Sensitivity to system parameters.ParameterModelExecution time (s)SLA violation rate (%)Workload IntensityDQL28934(500 concurrent tasks)DQL-PSO18612Number of VMsPSO24127(from 30 to 60)DQL-PSO154 → 14223 → 11Host CapacityGA27230(from 8000 to 16000 MIPS)DQL-PSO201 → 15919 → 8 Table 8Sensitivity to system parameters.ParameterModelExecution time (s)SLA violation rate (%)Workload IntensityDQL28934(500 concurrent tasks)DQL-PSO18612Number of VMsPSO24127(from 30 to 60)DQL-PSO154 → 14223 → 11Host CapacityGA27230(from 8000 to 16000 MIPS)DQL-PSO201 → 15919 → 8 Sensitivity to system parameters. Conclusion The results of this study demonstrate that the proposed hybrid algorithm combining Deep Q-Learning (DQL) and Particle Swarm Optimization (PSO) significantly improves resource allocation efficiency in cloud computing environments. The integration of reinforcement learning with swarm intelligence enables the system to adapt dynamically to changing workload conditions while maintaining high resource utilization and minimizing SLA violations. Compared to standalone DQL and PSO approaches, as well as other hybrid and heuristic-based methods, the proposed model achieved superior performance across multiple metrics, including task execution time, energy consumption, and convergence speed. These improvements can be attributed to the complementary strengths of both algorithms: DQL provides long-term policy adaptation through environmental interaction, while PSO enhances action selection by performing efficient global search optimization. This synergy allows the system to balance exploration and exploitation effectively, leading to more stable and optimal decisions in real-time scenarios. Moreover, statistical analysis confirmed that the observed performance gains were not due to chance, reinforcing the robustness and reliability of the proposed method. Sensitivity tests further illustrated the algorithm’s resilience under varying system parameters such as workload intensity and VM availability, highlighting its practical applicability in real-world cloud infrastructures. However, the model does come with certain limitations. Its computational complexity is relatively higher than that of simpler heuristic strategies, which may affect deployment in latency-sensitive applications. Additionally, the effectiveness of the learning process depends heavily on the quality and representativeness of training data, suggesting that continuous adaptation mechanisms or online learning components could be beneficial for long-term operation. Despite these challenges, the proposed hybrid DQL-PSO framework represents a significant step forward in intelligent cloud resource management. It provides a scalable and adaptive solution that outperforms traditional and existing hybrid approaches, offering valuable insights into how machine learning and metaheuristic optimization can be combined to address complex scheduling problems. Future work may explore extending this approach to edge and fog computing environments, incorporating multi-agent coordination, or integrating federated learning for decentralized decision-making across distributed cloud systems. The hybrid design has moderate training overhead and shows sensitivity to reward weights and PSO hyperparameters. Although estimation is real-time in our settings, large-scale multi-tenant deployments may require further optimization. Future work will explore multi-agent reinforcement learning for cross-data center coordination, federated learning to enable decentralized training without raw data sharing, online hyperparameter optimization to reduce tuning effort, and extensions to hybrid cloud/edge scenarios with heterogeneous latency and energy profiles.